Commit | Line | Data |
---|---|---|
2e04ef76 RR |
1 | /*P:700 |
2 | * The pagetable code, on the other hand, still shows the scars of | |
f938d2c8 RR |
3 | * previous encounters. It's functional, and as neat as it can be in the |
4 | * circumstances, but be wary, for these things are subtle and break easily. | |
5 | * The Guest provides a virtual to physical mapping, but we can neither trust | |
a6bd8e13 | 6 | * it nor use it: we verify and convert it here then point the CPU to the |
2e04ef76 RR |
7 | * converted Guest pages when running the Guest. |
8 | :*/ | |
f938d2c8 RR |
9 | |
10 | /* Copyright (C) Rusty Russell IBM Corporation 2006. | |
d7e28ffe RR |
11 | * GPL v2 and any later version */ |
12 | #include <linux/mm.h> | |
5a0e3ad6 | 13 | #include <linux/gfp.h> |
d7e28ffe RR |
14 | #include <linux/types.h> |
15 | #include <linux/spinlock.h> | |
16 | #include <linux/random.h> | |
17 | #include <linux/percpu.h> | |
18 | #include <asm/tlbflush.h> | |
47436aa4 | 19 | #include <asm/uaccess.h> |
d7e28ffe RR |
20 | #include "lg.h" |
21 | ||
2e04ef76 RR |
22 | /*M:008 |
23 | * We hold reference to pages, which prevents them from being swapped. | |
f56a384e RR |
24 | * It'd be nice to have a callback in the "struct mm_struct" when Linux wants |
25 | * to swap out. If we had this, and a shrinker callback to trim PTE pages, we | |
2e04ef76 RR |
26 | * could probably consider launching Guests as non-root. |
27 | :*/ | |
f56a384e | 28 | |
bff672e6 RR |
29 | /*H:300 |
30 | * The Page Table Code | |
31 | * | |
a91d74a3 RR |
32 | * We use two-level page tables for the Guest, or three-level with PAE. If |
33 | * you're not entirely comfortable with virtual addresses, physical addresses | |
34 | * and page tables then I recommend you review arch/x86/lguest/boot.c's "Page | |
35 | * Table Handling" (with diagrams!). | |
bff672e6 RR |
36 | * |
37 | * The Guest keeps page tables, but we maintain the actual ones here: these are | |
38 | * called "shadow" page tables. Which is a very Guest-centric name: these are | |
39 | * the real page tables the CPU uses, although we keep them up to date to | |
40 | * reflect the Guest's. (See what I mean about weird naming? Since when do | |
41 | * shadows reflect anything?) | |
42 | * | |
43 | * Anyway, this is the most complicated part of the Host code. There are seven | |
44 | * parts to this: | |
e1e72965 RR |
45 | * (i) Looking up a page table entry when the Guest faults, |
46 | * (ii) Making sure the Guest stack is mapped, | |
47 | * (iii) Setting up a page table entry when the Guest tells us one has changed, | |
bff672e6 | 48 | * (iv) Switching page tables, |
e1e72965 | 49 | * (v) Flushing (throwing away) page tables, |
bff672e6 RR |
50 | * (vi) Mapping the Switcher when the Guest is about to run, |
51 | * (vii) Setting up the page tables initially. | |
2e04ef76 | 52 | :*/ |
bff672e6 | 53 | |
2e04ef76 | 54 | /* |
a91d74a3 RR |
55 | * The Switcher uses the complete top PTE page. That's 1024 PTE entries (4MB) |
56 | * or 512 PTE entries with PAE (2MB). | |
2e04ef76 | 57 | */ |
df29f43e | 58 | #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1) |
d7e28ffe | 59 | |
2e04ef76 RR |
60 | /* |
61 | * For PAE we need the PMD index as well. We use the last 2MB, so we | |
62 | * will need the last pmd entry of the last pmd page. | |
63 | */ | |
acdd0b62 MZ |
64 | #ifdef CONFIG_X86_PAE |
65 | #define SWITCHER_PMD_INDEX (PTRS_PER_PMD - 1) | |
66 | #define RESERVE_MEM 2U | |
67 | #define CHECK_GPGD_MASK _PAGE_PRESENT | |
68 | #else | |
69 | #define RESERVE_MEM 4U | |
70 | #define CHECK_GPGD_MASK _PAGE_TABLE | |
71 | #endif | |
72 | ||
2e04ef76 RR |
73 | /* |
74 | * We actually need a separate PTE page for each CPU. Remember that after the | |
bff672e6 | 75 | * Switcher code itself comes two pages for each CPU, and we don't want this |
2e04ef76 RR |
76 | * CPU's guest to see the pages of any other CPU. |
77 | */ | |
df29f43e | 78 | static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); |
d7e28ffe RR |
79 | #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) |
80 | ||
2e04ef76 RR |
81 | /*H:320 |
82 | * The page table code is curly enough to need helper functions to keep it | |
a91d74a3 RR |
83 | * clear and clean. The kernel itself provides many of them; one advantage |
84 | * of insisting that the Guest and Host use the same CONFIG_PAE setting. | |
bff672e6 | 85 | * |
df29f43e | 86 | * There are two functions which return pointers to the shadow (aka "real") |
bff672e6 RR |
87 | * page tables. |
88 | * | |
89 | * spgd_addr() takes the virtual address and returns a pointer to the top-level | |
e1e72965 RR |
90 | * page directory entry (PGD) for that address. Since we keep track of several |
91 | * page tables, the "i" argument tells us which one we're interested in (it's | |
2e04ef76 RR |
92 | * usually the current one). |
93 | */ | |
382ac6b3 | 94 | static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr) |
d7e28ffe | 95 | { |
df29f43e | 96 | unsigned int index = pgd_index(vaddr); |
d7e28ffe | 97 | |
bff672e6 | 98 | /* Return a pointer index'th pgd entry for the i'th page table. */ |
382ac6b3 | 99 | return &cpu->lg->pgdirs[i].pgdir[index]; |
d7e28ffe RR |
100 | } |
101 | ||
acdd0b62 | 102 | #ifdef CONFIG_X86_PAE |
2e04ef76 RR |
103 | /* |
104 | * This routine then takes the PGD entry given above, which contains the | |
acdd0b62 | 105 | * address of the PMD page. It then returns a pointer to the PMD entry for the |
2e04ef76 RR |
106 | * given address. |
107 | */ | |
acdd0b62 MZ |
108 | static pmd_t *spmd_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) |
109 | { | |
110 | unsigned int index = pmd_index(vaddr); | |
111 | pmd_t *page; | |
112 | ||
acdd0b62 MZ |
113 | /* You should never call this if the PGD entry wasn't valid */ |
114 | BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); | |
115 | page = __va(pgd_pfn(spgd) << PAGE_SHIFT); | |
116 | ||
117 | return &page[index]; | |
118 | } | |
119 | #endif | |
120 | ||
2e04ef76 RR |
121 | /* |
122 | * This routine then takes the page directory entry returned above, which | |
e1e72965 | 123 | * contains the address of the page table entry (PTE) page. It then returns a |
2e04ef76 RR |
124 | * pointer to the PTE entry for the given address. |
125 | */ | |
acdd0b62 | 126 | static pte_t *spte_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) |
d7e28ffe | 127 | { |
acdd0b62 MZ |
128 | #ifdef CONFIG_X86_PAE |
129 | pmd_t *pmd = spmd_addr(cpu, spgd, vaddr); | |
130 | pte_t *page = __va(pmd_pfn(*pmd) << PAGE_SHIFT); | |
131 | ||
132 | /* You should never call this if the PMD entry wasn't valid */ | |
133 | BUG_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)); | |
134 | #else | |
df29f43e | 135 | pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT); |
bff672e6 | 136 | /* You should never call this if the PGD entry wasn't valid */ |
df29f43e | 137 | BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); |
acdd0b62 MZ |
138 | #endif |
139 | ||
90603d15 | 140 | return &page[pte_index(vaddr)]; |
d7e28ffe RR |
141 | } |
142 | ||
2e04ef76 | 143 | /* |
9f54288d | 144 | * These functions are just like the above, except they access the Guest |
2e04ef76 RR |
145 | * page tables. Hence they return a Guest address. |
146 | */ | |
1713608f | 147 | static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 148 | { |
df29f43e | 149 | unsigned int index = vaddr >> (PGDIR_SHIFT); |
1713608f | 150 | return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t); |
d7e28ffe RR |
151 | } |
152 | ||
acdd0b62 | 153 | #ifdef CONFIG_X86_PAE |
a91d74a3 | 154 | /* Follow the PGD to the PMD. */ |
acdd0b62 | 155 | static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr) |
d7e28ffe | 156 | { |
df29f43e MZ |
157 | unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; |
158 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); | |
acdd0b62 MZ |
159 | return gpage + pmd_index(vaddr) * sizeof(pmd_t); |
160 | } | |
acdd0b62 | 161 | |
a91d74a3 | 162 | /* Follow the PMD to the PTE. */ |
acdd0b62 | 163 | static unsigned long gpte_addr(struct lg_cpu *cpu, |
92b4d8df | 164 | pmd_t gpmd, unsigned long vaddr) |
acdd0b62 | 165 | { |
92b4d8df | 166 | unsigned long gpage = pmd_pfn(gpmd) << PAGE_SHIFT; |
acdd0b62 | 167 | |
acdd0b62 | 168 | BUG_ON(!(pmd_flags(gpmd) & _PAGE_PRESENT)); |
92b4d8df RR |
169 | return gpage + pte_index(vaddr) * sizeof(pte_t); |
170 | } | |
acdd0b62 | 171 | #else |
a91d74a3 | 172 | /* Follow the PGD to the PTE (no mid-level for !PAE). */ |
92b4d8df RR |
173 | static unsigned long gpte_addr(struct lg_cpu *cpu, |
174 | pgd_t gpgd, unsigned long vaddr) | |
175 | { | |
176 | unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; | |
177 | ||
178 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); | |
90603d15 | 179 | return gpage + pte_index(vaddr) * sizeof(pte_t); |
d7e28ffe | 180 | } |
92b4d8df | 181 | #endif |
a6bd8e13 RR |
182 | /*:*/ |
183 | ||
9f54288d | 184 | /*M:007 |
2e04ef76 RR |
185 | * get_pfn is slow: we could probably try to grab batches of pages here as |
186 | * an optimization (ie. pre-faulting). | |
187 | :*/ | |
d7e28ffe | 188 | |
2e04ef76 RR |
189 | /*H:350 |
190 | * This routine takes a page number given by the Guest and converts it to | |
bff672e6 RR |
191 | * an actual, physical page number. It can fail for several reasons: the |
192 | * virtual address might not be mapped by the Launcher, the write flag is set | |
193 | * and the page is read-only, or the write flag was set and the page was | |
194 | * shared so had to be copied, but we ran out of memory. | |
195 | * | |
a6bd8e13 | 196 | * This holds a reference to the page, so release_pte() is careful to put that |
2e04ef76 RR |
197 | * back. |
198 | */ | |
d7e28ffe RR |
199 | static unsigned long get_pfn(unsigned long virtpfn, int write) |
200 | { | |
201 | struct page *page; | |
71a3f4ed RR |
202 | |
203 | /* gup me one page at this address please! */ | |
204 | if (get_user_pages_fast(virtpfn << PAGE_SHIFT, 1, write, &page) == 1) | |
205 | return page_to_pfn(page); | |
206 | ||
bff672e6 | 207 | /* This value indicates failure. */ |
71a3f4ed | 208 | return -1UL; |
d7e28ffe RR |
209 | } |
210 | ||
2e04ef76 RR |
211 | /*H:340 |
212 | * Converting a Guest page table entry to a shadow (ie. real) page table | |
bff672e6 RR |
213 | * entry can be a little tricky. The flags are (almost) the same, but the |
214 | * Guest PTE contains a virtual page number: the CPU needs the real page | |
2e04ef76 RR |
215 | * number. |
216 | */ | |
382ac6b3 | 217 | static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write) |
d7e28ffe | 218 | { |
df29f43e | 219 | unsigned long pfn, base, flags; |
d7e28ffe | 220 | |
2e04ef76 RR |
221 | /* |
222 | * The Guest sets the global flag, because it thinks that it is using | |
bff672e6 RR |
223 | * PGE. We only told it to use PGE so it would tell us whether it was |
224 | * flushing a kernel mapping or a userspace mapping. We don't actually | |
2e04ef76 RR |
225 | * use the global bit, so throw it away. |
226 | */ | |
df29f43e | 227 | flags = (pte_flags(gpte) & ~_PAGE_GLOBAL); |
bff672e6 | 228 | |
3c6b5bfa | 229 | /* The Guest's pages are offset inside the Launcher. */ |
382ac6b3 | 230 | base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE; |
3c6b5bfa | 231 | |
2e04ef76 RR |
232 | /* |
233 | * We need a temporary "unsigned long" variable to hold the answer from | |
bff672e6 RR |
234 | * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't |
235 | * fit in spte.pfn. get_pfn() finds the real physical number of the | |
2e04ef76 RR |
236 | * page, given the virtual number. |
237 | */ | |
df29f43e | 238 | pfn = get_pfn(base + pte_pfn(gpte), write); |
d7e28ffe | 239 | if (pfn == -1UL) { |
382ac6b3 | 240 | kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte)); |
2e04ef76 RR |
241 | /* |
242 | * When we destroy the Guest, we'll go through the shadow page | |
bff672e6 | 243 | * tables and release_pte() them. Make sure we don't think |
2e04ef76 RR |
244 | * this one is valid! |
245 | */ | |
df29f43e | 246 | flags = 0; |
d7e28ffe | 247 | } |
df29f43e MZ |
248 | /* Now we assemble our shadow PTE from the page number and flags. */ |
249 | return pfn_pte(pfn, __pgprot(flags)); | |
d7e28ffe RR |
250 | } |
251 | ||
bff672e6 | 252 | /*H:460 And to complete the chain, release_pte() looks like this: */ |
df29f43e | 253 | static void release_pte(pte_t pte) |
d7e28ffe | 254 | { |
2e04ef76 RR |
255 | /* |
256 | * Remember that get_user_pages_fast() took a reference to the page, in | |
257 | * get_pfn()? We have to put it back now. | |
258 | */ | |
df29f43e | 259 | if (pte_flags(pte) & _PAGE_PRESENT) |
90603d15 | 260 | put_page(pte_page(pte)); |
d7e28ffe | 261 | } |
bff672e6 | 262 | /*:*/ |
d7e28ffe | 263 | |
382ac6b3 | 264 | static void check_gpte(struct lg_cpu *cpu, pte_t gpte) |
d7e28ffe | 265 | { |
31f4b46e AD |
266 | if ((pte_flags(gpte) & _PAGE_PSE) || |
267 | pte_pfn(gpte) >= cpu->lg->pfn_limit) | |
382ac6b3 | 268 | kill_guest(cpu, "bad page table entry"); |
d7e28ffe RR |
269 | } |
270 | ||
382ac6b3 | 271 | static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd) |
d7e28ffe | 272 | { |
acdd0b62 | 273 | if ((pgd_flags(gpgd) & ~CHECK_GPGD_MASK) || |
382ac6b3 GOC |
274 | (pgd_pfn(gpgd) >= cpu->lg->pfn_limit)) |
275 | kill_guest(cpu, "bad page directory entry"); | |
d7e28ffe RR |
276 | } |
277 | ||
acdd0b62 MZ |
278 | #ifdef CONFIG_X86_PAE |
279 | static void check_gpmd(struct lg_cpu *cpu, pmd_t gpmd) | |
280 | { | |
281 | if ((pmd_flags(gpmd) & ~_PAGE_TABLE) || | |
282 | (pmd_pfn(gpmd) >= cpu->lg->pfn_limit)) | |
283 | kill_guest(cpu, "bad page middle directory entry"); | |
284 | } | |
285 | #endif | |
286 | ||
bff672e6 | 287 | /*H:330 |
e1e72965 | 288 | * (i) Looking up a page table entry when the Guest faults. |
bff672e6 RR |
289 | * |
290 | * We saw this call in run_guest(): when we see a page fault in the Guest, we | |
291 | * come here. That's because we only set up the shadow page tables lazily as | |
292 | * they're needed, so we get page faults all the time and quietly fix them up | |
293 | * and return to the Guest without it knowing. | |
294 | * | |
295 | * If we fixed up the fault (ie. we mapped the address), this routine returns | |
2e04ef76 RR |
296 | * true. Otherwise, it was a real fault and we need to tell the Guest. |
297 | */ | |
df1693ab | 298 | bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) |
d7e28ffe | 299 | { |
df29f43e MZ |
300 | pgd_t gpgd; |
301 | pgd_t *spgd; | |
d7e28ffe | 302 | unsigned long gpte_ptr; |
df29f43e MZ |
303 | pte_t gpte; |
304 | pte_t *spte; | |
d7e28ffe | 305 | |
a91d74a3 | 306 | /* Mid level for PAE. */ |
acdd0b62 MZ |
307 | #ifdef CONFIG_X86_PAE |
308 | pmd_t *spmd; | |
309 | pmd_t gpmd; | |
310 | #endif | |
311 | ||
68a644d7 RR |
312 | /* We never demand page the Switcher, so trying is a mistake. */ |
313 | if (vaddr >= switcher_addr) | |
314 | return false; | |
315 | ||
bff672e6 | 316 | /* First step: get the top-level Guest page table entry. */ |
5dea1c88 RR |
317 | if (unlikely(cpu->linear_pages)) { |
318 | /* Faking up a linear mapping. */ | |
319 | gpgd = __pgd(CHECK_GPGD_MASK); | |
320 | } else { | |
321 | gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); | |
322 | /* Toplevel not present? We can't map it in. */ | |
323 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) | |
324 | return false; | |
325 | } | |
d7e28ffe | 326 | |
bff672e6 | 327 | /* Now look at the matching shadow entry. */ |
382ac6b3 | 328 | spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); |
df29f43e | 329 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) { |
bff672e6 | 330 | /* No shadow entry: allocate a new shadow PTE page. */ |
d7e28ffe | 331 | unsigned long ptepage = get_zeroed_page(GFP_KERNEL); |
2e04ef76 RR |
332 | /* |
333 | * This is not really the Guest's fault, but killing it is | |
334 | * simple for this corner case. | |
335 | */ | |
d7e28ffe | 336 | if (!ptepage) { |
382ac6b3 | 337 | kill_guest(cpu, "out of memory allocating pte page"); |
df1693ab | 338 | return false; |
d7e28ffe | 339 | } |
bff672e6 | 340 | /* We check that the Guest pgd is OK. */ |
382ac6b3 | 341 | check_gpgd(cpu, gpgd); |
2e04ef76 RR |
342 | /* |
343 | * And we copy the flags to the shadow PGD entry. The page | |
344 | * number in the shadow PGD is the page we just allocated. | |
345 | */ | |
acdd0b62 | 346 | set_pgd(spgd, __pgd(__pa(ptepage) | pgd_flags(gpgd))); |
d7e28ffe RR |
347 | } |
348 | ||
acdd0b62 | 349 | #ifdef CONFIG_X86_PAE |
5dea1c88 RR |
350 | if (unlikely(cpu->linear_pages)) { |
351 | /* Faking up a linear mapping. */ | |
352 | gpmd = __pmd(_PAGE_TABLE); | |
353 | } else { | |
354 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); | |
355 | /* Middle level not present? We can't map it in. */ | |
356 | if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) | |
357 | return false; | |
358 | } | |
acdd0b62 MZ |
359 | |
360 | /* Now look at the matching shadow entry. */ | |
361 | spmd = spmd_addr(cpu, *spgd, vaddr); | |
362 | ||
363 | if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) { | |
364 | /* No shadow entry: allocate a new shadow PTE page. */ | |
365 | unsigned long ptepage = get_zeroed_page(GFP_KERNEL); | |
366 | ||
2e04ef76 RR |
367 | /* |
368 | * This is not really the Guest's fault, but killing it is | |
369 | * simple for this corner case. | |
370 | */ | |
acdd0b62 MZ |
371 | if (!ptepage) { |
372 | kill_guest(cpu, "out of memory allocating pte page"); | |
373 | return false; | |
374 | } | |
375 | ||
376 | /* We check that the Guest pmd is OK. */ | |
377 | check_gpmd(cpu, gpmd); | |
378 | ||
2e04ef76 RR |
379 | /* |
380 | * And we copy the flags to the shadow PMD entry. The page | |
381 | * number in the shadow PMD is the page we just allocated. | |
382 | */ | |
4c1ea3dd | 383 | set_pmd(spmd, __pmd(__pa(ptepage) | pmd_flags(gpmd))); |
acdd0b62 | 384 | } |
92b4d8df | 385 | |
2e04ef76 RR |
386 | /* |
387 | * OK, now we look at the lower level in the Guest page table: keep its | |
388 | * address, because we might update it later. | |
389 | */ | |
92b4d8df RR |
390 | gpte_ptr = gpte_addr(cpu, gpmd, vaddr); |
391 | #else | |
2e04ef76 RR |
392 | /* |
393 | * OK, now we look at the lower level in the Guest page table: keep its | |
394 | * address, because we might update it later. | |
395 | */ | |
acdd0b62 | 396 | gpte_ptr = gpte_addr(cpu, gpgd, vaddr); |
92b4d8df | 397 | #endif |
a91d74a3 | 398 | |
5dea1c88 RR |
399 | if (unlikely(cpu->linear_pages)) { |
400 | /* Linear? Make up a PTE which points to same page. */ | |
401 | gpte = __pte((vaddr & PAGE_MASK) | _PAGE_RW | _PAGE_PRESENT); | |
402 | } else { | |
403 | /* Read the actual PTE value. */ | |
404 | gpte = lgread(cpu, gpte_ptr, pte_t); | |
405 | } | |
d7e28ffe | 406 | |
bff672e6 | 407 | /* If this page isn't in the Guest page tables, we can't page it in. */ |
df29f43e | 408 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
df1693ab | 409 | return false; |
d7e28ffe | 410 | |
2e04ef76 RR |
411 | /* |
412 | * Check they're not trying to write to a page the Guest wants | |
413 | * read-only (bit 2 of errcode == write). | |
414 | */ | |
df29f43e | 415 | if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW)) |
df1693ab | 416 | return false; |
d7e28ffe | 417 | |
e1e72965 | 418 | /* User access to a kernel-only page? (bit 3 == user access) */ |
df29f43e | 419 | if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER)) |
df1693ab | 420 | return false; |
d7e28ffe | 421 | |
2e04ef76 RR |
422 | /* |
423 | * Check that the Guest PTE flags are OK, and the page number is below | |
424 | * the pfn_limit (ie. not mapping the Launcher binary). | |
425 | */ | |
382ac6b3 | 426 | check_gpte(cpu, gpte); |
e1e72965 | 427 | |
bff672e6 | 428 | /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */ |
df29f43e | 429 | gpte = pte_mkyoung(gpte); |
d7e28ffe | 430 | if (errcode & 2) |
df29f43e | 431 | gpte = pte_mkdirty(gpte); |
d7e28ffe | 432 | |
bff672e6 | 433 | /* Get the pointer to the shadow PTE entry we're going to set. */ |
acdd0b62 | 434 | spte = spte_addr(cpu, *spgd, vaddr); |
2e04ef76 RR |
435 | |
436 | /* | |
437 | * If there was a valid shadow PTE entry here before, we release it. | |
438 | * This can happen with a write to a previously read-only entry. | |
439 | */ | |
d7e28ffe RR |
440 | release_pte(*spte); |
441 | ||
2e04ef76 RR |
442 | /* |
443 | * If this is a write, we insist that the Guest page is writable (the | |
444 | * final arg to gpte_to_spte()). | |
445 | */ | |
df29f43e | 446 | if (pte_dirty(gpte)) |
382ac6b3 | 447 | *spte = gpte_to_spte(cpu, gpte, 1); |
df29f43e | 448 | else |
2e04ef76 RR |
449 | /* |
450 | * If this is a read, don't set the "writable" bit in the page | |
bff672e6 | 451 | * table entry, even if the Guest says it's writable. That way |
e1e72965 | 452 | * we will come back here when a write does actually occur, so |
2e04ef76 RR |
453 | * we can update the Guest's _PAGE_DIRTY flag. |
454 | */ | |
4c1ea3dd | 455 | set_pte(spte, gpte_to_spte(cpu, pte_wrprotect(gpte), 0)); |
d7e28ffe | 456 | |
2e04ef76 RR |
457 | /* |
458 | * Finally, we write the Guest PTE entry back: we've set the | |
459 | * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. | |
460 | */ | |
5dea1c88 RR |
461 | if (likely(!cpu->linear_pages)) |
462 | lgwrite(cpu, gpte_ptr, pte_t, gpte); | |
bff672e6 | 463 | |
2e04ef76 RR |
464 | /* |
465 | * The fault is fixed, the page table is populated, the mapping | |
e1e72965 RR |
466 | * manipulated, the result returned and the code complete. A small |
467 | * delay and a trace of alliteration are the only indications the Guest | |
2e04ef76 RR |
468 | * has that a page fault occurred at all. |
469 | */ | |
df1693ab | 470 | return true; |
d7e28ffe RR |
471 | } |
472 | ||
e1e72965 RR |
473 | /*H:360 |
474 | * (ii) Making sure the Guest stack is mapped. | |
bff672e6 | 475 | * |
e1e72965 RR |
476 | * Remember that direct traps into the Guest need a mapped Guest kernel stack. |
477 | * pin_stack_pages() calls us here: we could simply call demand_page(), but as | |
478 | * we've seen that logic is quite long, and usually the stack pages are already | |
479 | * mapped, so it's overkill. | |
bff672e6 RR |
480 | * |
481 | * This is a quick version which answers the question: is this virtual address | |
2e04ef76 RR |
482 | * mapped by the shadow page tables, and is it writable? |
483 | */ | |
df1693ab | 484 | static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 485 | { |
df29f43e | 486 | pgd_t *spgd; |
d7e28ffe | 487 | unsigned long flags; |
acdd0b62 MZ |
488 | #ifdef CONFIG_X86_PAE |
489 | pmd_t *spmd; | |
490 | #endif | |
68a644d7 RR |
491 | |
492 | /* You can't put your stack in the Switcher! */ | |
493 | if (vaddr >= switcher_addr) | |
494 | return false; | |
495 | ||
e1e72965 | 496 | /* Look at the current top level entry: is it present? */ |
382ac6b3 | 497 | spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); |
df29f43e | 498 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) |
df1693ab | 499 | return false; |
d7e28ffe | 500 | |
acdd0b62 MZ |
501 | #ifdef CONFIG_X86_PAE |
502 | spmd = spmd_addr(cpu, *spgd, vaddr); | |
503 | if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) | |
504 | return false; | |
505 | #endif | |
506 | ||
2e04ef76 RR |
507 | /* |
508 | * Check the flags on the pte entry itself: it must be present and | |
509 | * writable. | |
510 | */ | |
acdd0b62 | 511 | flags = pte_flags(*(spte_addr(cpu, *spgd, vaddr))); |
df29f43e | 512 | |
d7e28ffe RR |
513 | return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); |
514 | } | |
515 | ||
2e04ef76 RR |
516 | /* |
517 | * So, when pin_stack_pages() asks us to pin a page, we check if it's already | |
bff672e6 | 518 | * in the page tables, and if not, we call demand_page() with error code 2 |
2e04ef76 RR |
519 | * (meaning "write"). |
520 | */ | |
1713608f | 521 | void pin_page(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 522 | { |
1713608f | 523 | if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2)) |
382ac6b3 | 524 | kill_guest(cpu, "bad stack page %#lx", vaddr); |
d7e28ffe | 525 | } |
a91d74a3 | 526 | /*:*/ |
d7e28ffe | 527 | |
acdd0b62 MZ |
528 | #ifdef CONFIG_X86_PAE |
529 | static void release_pmd(pmd_t *spmd) | |
530 | { | |
531 | /* If the entry's not present, there's nothing to release. */ | |
532 | if (pmd_flags(*spmd) & _PAGE_PRESENT) { | |
533 | unsigned int i; | |
534 | pte_t *ptepage = __va(pmd_pfn(*spmd) << PAGE_SHIFT); | |
535 | /* For each entry in the page, we might need to release it. */ | |
536 | for (i = 0; i < PTRS_PER_PTE; i++) | |
537 | release_pte(ptepage[i]); | |
538 | /* Now we can free the page of PTEs */ | |
539 | free_page((long)ptepage); | |
540 | /* And zero out the PMD entry so we never release it twice. */ | |
4c1ea3dd | 541 | set_pmd(spmd, __pmd(0)); |
acdd0b62 MZ |
542 | } |
543 | } | |
544 | ||
545 | static void release_pgd(pgd_t *spgd) | |
546 | { | |
547 | /* If the entry's not present, there's nothing to release. */ | |
548 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { | |
549 | unsigned int i; | |
550 | pmd_t *pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); | |
551 | ||
552 | for (i = 0; i < PTRS_PER_PMD; i++) | |
553 | release_pmd(&pmdpage[i]); | |
554 | ||
555 | /* Now we can free the page of PMDs */ | |
556 | free_page((long)pmdpage); | |
557 | /* And zero out the PGD entry so we never release it twice. */ | |
558 | set_pgd(spgd, __pgd(0)); | |
559 | } | |
560 | } | |
561 | ||
562 | #else /* !CONFIG_X86_PAE */ | |
a91d74a3 RR |
563 | /*H:450 |
564 | * If we chase down the release_pgd() code, the non-PAE version looks like | |
565 | * this. The PAE version is almost identical, but instead of calling | |
566 | * release_pte it calls release_pmd(), which looks much like this. | |
567 | */ | |
90603d15 | 568 | static void release_pgd(pgd_t *spgd) |
d7e28ffe | 569 | { |
bff672e6 | 570 | /* If the entry's not present, there's nothing to release. */ |
df29f43e | 571 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
d7e28ffe | 572 | unsigned int i; |
2e04ef76 RR |
573 | /* |
574 | * Converting the pfn to find the actual PTE page is easy: turn | |
bff672e6 | 575 | * the page number into a physical address, then convert to a |
2e04ef76 RR |
576 | * virtual address (easy for kernel pages like this one). |
577 | */ | |
df29f43e | 578 | pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); |
bff672e6 | 579 | /* For each entry in the page, we might need to release it. */ |
df29f43e | 580 | for (i = 0; i < PTRS_PER_PTE; i++) |
d7e28ffe | 581 | release_pte(ptepage[i]); |
bff672e6 | 582 | /* Now we can free the page of PTEs */ |
d7e28ffe | 583 | free_page((long)ptepage); |
e1e72965 | 584 | /* And zero out the PGD entry so we never release it twice. */ |
df29f43e | 585 | *spgd = __pgd(0); |
d7e28ffe RR |
586 | } |
587 | } | |
acdd0b62 | 588 | #endif |
2e04ef76 RR |
589 | |
590 | /*H:445 | |
591 | * We saw flush_user_mappings() twice: once from the flush_user_mappings() | |
e1e72965 | 592 | * hypercall and once in new_pgdir() when we re-used a top-level pgdir page. |
2e04ef76 RR |
593 | * It simply releases every PTE page from 0 up to the Guest's kernel address. |
594 | */ | |
d7e28ffe RR |
595 | static void flush_user_mappings(struct lguest *lg, int idx) |
596 | { | |
597 | unsigned int i; | |
bff672e6 | 598 | /* Release every pgd entry up to the kernel's address. */ |
47436aa4 | 599 | for (i = 0; i < pgd_index(lg->kernel_address); i++) |
90603d15 | 600 | release_pgd(lg->pgdirs[idx].pgdir + i); |
d7e28ffe RR |
601 | } |
602 | ||
2e04ef76 RR |
603 | /*H:440 |
604 | * (v) Flushing (throwing away) page tables, | |
e1e72965 RR |
605 | * |
606 | * The Guest has a hypercall to throw away the page tables: it's used when a | |
2e04ef76 RR |
607 | * large number of mappings have been changed. |
608 | */ | |
1713608f | 609 | void guest_pagetable_flush_user(struct lg_cpu *cpu) |
d7e28ffe | 610 | { |
bff672e6 | 611 | /* Drop the userspace part of the current page table. */ |
1713608f | 612 | flush_user_mappings(cpu->lg, cpu->cpu_pgd); |
d7e28ffe | 613 | } |
bff672e6 | 614 | /*:*/ |
d7e28ffe | 615 | |
47436aa4 | 616 | /* We walk down the guest page tables to get a guest-physical address */ |
1713608f | 617 | unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr) |
47436aa4 RR |
618 | { |
619 | pgd_t gpgd; | |
620 | pte_t gpte; | |
acdd0b62 MZ |
621 | #ifdef CONFIG_X86_PAE |
622 | pmd_t gpmd; | |
623 | #endif | |
5dea1c88 RR |
624 | |
625 | /* Still not set up? Just map 1:1. */ | |
626 | if (unlikely(cpu->linear_pages)) | |
627 | return vaddr; | |
628 | ||
47436aa4 | 629 | /* First step: get the top-level Guest page table entry. */ |
382ac6b3 | 630 | gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); |
47436aa4 | 631 | /* Toplevel not present? We can't map it in. */ |
6afbdd05 | 632 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) { |
382ac6b3 | 633 | kill_guest(cpu, "Bad address %#lx", vaddr); |
6afbdd05 RR |
634 | return -1UL; |
635 | } | |
47436aa4 | 636 | |
acdd0b62 MZ |
637 | #ifdef CONFIG_X86_PAE |
638 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); | |
639 | if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) | |
640 | kill_guest(cpu, "Bad address %#lx", vaddr); | |
92b4d8df RR |
641 | gpte = lgread(cpu, gpte_addr(cpu, gpmd, vaddr), pte_t); |
642 | #else | |
acdd0b62 | 643 | gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t); |
92b4d8df | 644 | #endif |
47436aa4 | 645 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
382ac6b3 | 646 | kill_guest(cpu, "Bad address %#lx", vaddr); |
47436aa4 RR |
647 | |
648 | return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK); | |
649 | } | |
650 | ||
2e04ef76 RR |
651 | /* |
652 | * We keep several page tables. This is a simple routine to find the page | |
bff672e6 | 653 | * table (if any) corresponding to this top-level address the Guest has given |
2e04ef76 RR |
654 | * us. |
655 | */ | |
d7e28ffe RR |
656 | static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) |
657 | { | |
658 | unsigned int i; | |
659 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) | |
4357bd94 | 660 | if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable) |
d7e28ffe RR |
661 | break; |
662 | return i; | |
663 | } | |
664 | ||
2e04ef76 RR |
665 | /*H:435 |
666 | * And this is us, creating the new page directory. If we really do | |
bff672e6 | 667 | * allocate a new one (and so the kernel parts are not there), we set |
2e04ef76 RR |
668 | * blank_pgdir. |
669 | */ | |
1713608f | 670 | static unsigned int new_pgdir(struct lg_cpu *cpu, |
ee3db0f2 | 671 | unsigned long gpgdir, |
d7e28ffe RR |
672 | int *blank_pgdir) |
673 | { | |
674 | unsigned int next; | |
acdd0b62 MZ |
675 | #ifdef CONFIG_X86_PAE |
676 | pmd_t *pmd_table; | |
677 | #endif | |
d7e28ffe | 678 | |
2e04ef76 RR |
679 | /* |
680 | * We pick one entry at random to throw out. Choosing the Least | |
681 | * Recently Used might be better, but this is easy. | |
682 | */ | |
382ac6b3 | 683 | next = random32() % ARRAY_SIZE(cpu->lg->pgdirs); |
bff672e6 | 684 | /* If it's never been allocated at all before, try now. */ |
382ac6b3 GOC |
685 | if (!cpu->lg->pgdirs[next].pgdir) { |
686 | cpu->lg->pgdirs[next].pgdir = | |
687 | (pgd_t *)get_zeroed_page(GFP_KERNEL); | |
bff672e6 | 688 | /* If the allocation fails, just keep using the one we have */ |
382ac6b3 | 689 | if (!cpu->lg->pgdirs[next].pgdir) |
1713608f | 690 | next = cpu->cpu_pgd; |
acdd0b62 MZ |
691 | else { |
692 | #ifdef CONFIG_X86_PAE | |
2e04ef76 RR |
693 | /* |
694 | * In PAE mode, allocate a pmd page and populate the | |
695 | * last pgd entry. | |
696 | */ | |
acdd0b62 MZ |
697 | pmd_table = (pmd_t *)get_zeroed_page(GFP_KERNEL); |
698 | if (!pmd_table) { | |
699 | free_page((long)cpu->lg->pgdirs[next].pgdir); | |
700 | set_pgd(cpu->lg->pgdirs[next].pgdir, __pgd(0)); | |
701 | next = cpu->cpu_pgd; | |
702 | } else { | |
703 | set_pgd(cpu->lg->pgdirs[next].pgdir + | |
704 | SWITCHER_PGD_INDEX, | |
705 | __pgd(__pa(pmd_table) | _PAGE_PRESENT)); | |
2e04ef76 RR |
706 | /* |
707 | * This is a blank page, so there are no kernel | |
708 | * mappings: caller must map the stack! | |
709 | */ | |
acdd0b62 MZ |
710 | *blank_pgdir = 1; |
711 | } | |
712 | #else | |
d7e28ffe | 713 | *blank_pgdir = 1; |
acdd0b62 MZ |
714 | #endif |
715 | } | |
d7e28ffe | 716 | } |
bff672e6 | 717 | /* Record which Guest toplevel this shadows. */ |
382ac6b3 | 718 | cpu->lg->pgdirs[next].gpgdir = gpgdir; |
d7e28ffe | 719 | /* Release all the non-kernel mappings. */ |
382ac6b3 | 720 | flush_user_mappings(cpu->lg, next); |
d7e28ffe RR |
721 | |
722 | return next; | |
723 | } | |
724 | ||
2e04ef76 RR |
725 | /*H:470 |
726 | * Finally, a routine which throws away everything: all PGD entries in all | |
e1e72965 | 727 | * the shadow page tables, including the Guest's kernel mappings. This is used |
2e04ef76 RR |
728 | * when we destroy the Guest. |
729 | */ | |
d7e28ffe RR |
730 | static void release_all_pagetables(struct lguest *lg) |
731 | { | |
732 | unsigned int i, j; | |
733 | ||
bff672e6 | 734 | /* Every shadow pagetable this Guest has */ |
d7e28ffe | 735 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
acdd0b62 MZ |
736 | if (lg->pgdirs[i].pgdir) { |
737 | #ifdef CONFIG_X86_PAE | |
738 | pgd_t *spgd; | |
739 | pmd_t *pmdpage; | |
740 | unsigned int k; | |
741 | ||
742 | /* Get the last pmd page. */ | |
743 | spgd = lg->pgdirs[i].pgdir + SWITCHER_PGD_INDEX; | |
744 | pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); | |
745 | ||
2e04ef76 RR |
746 | /* |
747 | * And release the pmd entries of that pmd page, | |
748 | * except for the switcher pmd. | |
749 | */ | |
acdd0b62 MZ |
750 | for (k = 0; k < SWITCHER_PMD_INDEX; k++) |
751 | release_pmd(&pmdpage[k]); | |
752 | #endif | |
bff672e6 | 753 | /* Every PGD entry except the Switcher at the top */ |
d7e28ffe | 754 | for (j = 0; j < SWITCHER_PGD_INDEX; j++) |
90603d15 | 755 | release_pgd(lg->pgdirs[i].pgdir + j); |
acdd0b62 | 756 | } |
d7e28ffe RR |
757 | } |
758 | ||
2e04ef76 RR |
759 | /* |
760 | * We also throw away everything when a Guest tells us it's changed a kernel | |
bff672e6 | 761 | * mapping. Since kernel mappings are in every page table, it's easiest to |
e1e72965 | 762 | * throw them all away. This traps the Guest in amber for a while as |
2e04ef76 RR |
763 | * everything faults back in, but it's rare. |
764 | */ | |
4665ac8e | 765 | void guest_pagetable_clear_all(struct lg_cpu *cpu) |
d7e28ffe | 766 | { |
4665ac8e | 767 | release_all_pagetables(cpu->lg); |
bff672e6 | 768 | /* We need the Guest kernel stack mapped again. */ |
4665ac8e | 769 | pin_stack_pages(cpu); |
d7e28ffe | 770 | } |
5dea1c88 RR |
771 | |
772 | /*H:430 | |
773 | * (iv) Switching page tables | |
774 | * | |
775 | * Now we've seen all the page table setting and manipulation, let's see | |
776 | * what happens when the Guest changes page tables (ie. changes the top-level | |
777 | * pgdir). This occurs on almost every context switch. | |
778 | */ | |
779 | void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable) | |
780 | { | |
781 | int newpgdir, repin = 0; | |
782 | ||
783 | /* | |
784 | * The very first time they call this, we're actually running without | |
785 | * any page tables; we've been making it up. Throw them away now. | |
786 | */ | |
787 | if (unlikely(cpu->linear_pages)) { | |
788 | release_all_pagetables(cpu->lg); | |
789 | cpu->linear_pages = false; | |
790 | /* Force allocation of a new pgdir. */ | |
791 | newpgdir = ARRAY_SIZE(cpu->lg->pgdirs); | |
792 | } else { | |
793 | /* Look to see if we have this one already. */ | |
794 | newpgdir = find_pgdir(cpu->lg, pgtable); | |
795 | } | |
796 | ||
797 | /* | |
798 | * If not, we allocate or mug an existing one: if it's a fresh one, | |
799 | * repin gets set to 1. | |
800 | */ | |
801 | if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs)) | |
802 | newpgdir = new_pgdir(cpu, pgtable, &repin); | |
803 | /* Change the current pgd index to the new one. */ | |
804 | cpu->cpu_pgd = newpgdir; | |
805 | /* If it was completely blank, we map in the Guest kernel stack */ | |
806 | if (repin) | |
807 | pin_stack_pages(cpu); | |
808 | } | |
e1e72965 | 809 | /*:*/ |
2e04ef76 RR |
810 | |
811 | /*M:009 | |
812 | * Since we throw away all mappings when a kernel mapping changes, our | |
e1e72965 RR |
813 | * performance sucks for guests using highmem. In fact, a guest with |
814 | * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is | |
815 | * usually slower than a Guest with less memory. | |
816 | * | |
817 | * This, of course, cannot be fixed. It would take some kind of... well, I | |
2e04ef76 RR |
818 | * don't know, but the term "puissant code-fu" comes to mind. |
819 | :*/ | |
d7e28ffe | 820 | |
2e04ef76 RR |
821 | /*H:420 |
822 | * This is the routine which actually sets the page table entry for then | |
bff672e6 RR |
823 | * "idx"'th shadow page table. |
824 | * | |
825 | * Normally, we can just throw out the old entry and replace it with 0: if they | |
826 | * use it demand_page() will put the new entry in. We need to do this anyway: | |
827 | * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page | |
828 | * is read from, and _PAGE_DIRTY when it's written to. | |
829 | * | |
830 | * But Avi Kivity pointed out that most Operating Systems (Linux included) set | |
831 | * these bits on PTEs immediately anyway. This is done to save the CPU from | |
832 | * having to update them, but it helps us the same way: if they set | |
833 | * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if | |
834 | * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately. | |
835 | */ | |
382ac6b3 | 836 | static void do_set_pte(struct lg_cpu *cpu, int idx, |
df29f43e | 837 | unsigned long vaddr, pte_t gpte) |
d7e28ffe | 838 | { |
e1e72965 | 839 | /* Look up the matching shadow page directory entry. */ |
382ac6b3 | 840 | pgd_t *spgd = spgd_addr(cpu, idx, vaddr); |
acdd0b62 MZ |
841 | #ifdef CONFIG_X86_PAE |
842 | pmd_t *spmd; | |
843 | #endif | |
bff672e6 RR |
844 | |
845 | /* If the top level isn't present, there's no entry to update. */ | |
df29f43e | 846 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
acdd0b62 MZ |
847 | #ifdef CONFIG_X86_PAE |
848 | spmd = spmd_addr(cpu, *spgd, vaddr); | |
849 | if (pmd_flags(*spmd) & _PAGE_PRESENT) { | |
850 | #endif | |
2e04ef76 | 851 | /* Otherwise, start by releasing the existing entry. */ |
acdd0b62 MZ |
852 | pte_t *spte = spte_addr(cpu, *spgd, vaddr); |
853 | release_pte(*spte); | |
854 | ||
2e04ef76 RR |
855 | /* |
856 | * If they're setting this entry as dirty or accessed, | |
857 | * we might as well put that entry they've given us in | |
858 | * now. This shaves 10% off a copy-on-write | |
859 | * micro-benchmark. | |
860 | */ | |
acdd0b62 MZ |
861 | if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) { |
862 | check_gpte(cpu, gpte); | |
4c1ea3dd RR |
863 | set_pte(spte, |
864 | gpte_to_spte(cpu, gpte, | |
acdd0b62 | 865 | pte_flags(gpte) & _PAGE_DIRTY)); |
2e04ef76 RR |
866 | } else { |
867 | /* | |
868 | * Otherwise kill it and we can demand_page() | |
869 | * it in later. | |
870 | */ | |
4c1ea3dd | 871 | set_pte(spte, __pte(0)); |
2e04ef76 | 872 | } |
acdd0b62 MZ |
873 | #ifdef CONFIG_X86_PAE |
874 | } | |
875 | #endif | |
d7e28ffe RR |
876 | } |
877 | } | |
878 | ||
2e04ef76 RR |
879 | /*H:410 |
880 | * Updating a PTE entry is a little trickier. | |
bff672e6 RR |
881 | * |
882 | * We keep track of several different page tables (the Guest uses one for each | |
883 | * process, so it makes sense to cache at least a few). Each of these have | |
884 | * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for | |
885 | * all processes. So when the page table above that address changes, we update | |
886 | * all the page tables, not just the current one. This is rare. | |
887 | * | |
a6bd8e13 | 888 | * The benefit is that when we have to track a new page table, we can keep all |
2e04ef76 RR |
889 | * the kernel mappings. This speeds up context switch immensely. |
890 | */ | |
382ac6b3 | 891 | void guest_set_pte(struct lg_cpu *cpu, |
ee3db0f2 | 892 | unsigned long gpgdir, unsigned long vaddr, pte_t gpte) |
d7e28ffe | 893 | { |
68a644d7 RR |
894 | /* We don't let you remap the Switcher; we need it to get back! */ |
895 | if (vaddr >= switcher_addr) { | |
896 | kill_guest(cpu, "attempt to set pte into Switcher pages"); | |
897 | return; | |
898 | } | |
899 | ||
2e04ef76 RR |
900 | /* |
901 | * Kernel mappings must be changed on all top levels. Slow, but doesn't | |
902 | * happen often. | |
903 | */ | |
382ac6b3 | 904 | if (vaddr >= cpu->lg->kernel_address) { |
d7e28ffe | 905 | unsigned int i; |
382ac6b3 GOC |
906 | for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++) |
907 | if (cpu->lg->pgdirs[i].pgdir) | |
908 | do_set_pte(cpu, i, vaddr, gpte); | |
d7e28ffe | 909 | } else { |
bff672e6 | 910 | /* Is this page table one we have a shadow for? */ |
382ac6b3 GOC |
911 | int pgdir = find_pgdir(cpu->lg, gpgdir); |
912 | if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs)) | |
bff672e6 | 913 | /* If so, do the update. */ |
382ac6b3 | 914 | do_set_pte(cpu, pgdir, vaddr, gpte); |
d7e28ffe RR |
915 | } |
916 | } | |
917 | ||
bff672e6 | 918 | /*H:400 |
e1e72965 | 919 | * (iii) Setting up a page table entry when the Guest tells us one has changed. |
bff672e6 RR |
920 | * |
921 | * Just like we did in interrupts_and_traps.c, it makes sense for us to deal | |
922 | * with the other side of page tables while we're here: what happens when the | |
923 | * Guest asks for a page table to be updated? | |
924 | * | |
925 | * We already saw that demand_page() will fill in the shadow page tables when | |
926 | * needed, so we can simply remove shadow page table entries whenever the Guest | |
927 | * tells us they've changed. When the Guest tries to use the new entry it will | |
928 | * fault and demand_page() will fix it up. | |
929 | * | |
fd589a8f | 930 | * So with that in mind here's our code to update a (top-level) PGD entry: |
bff672e6 | 931 | */ |
ebe0ba84 | 932 | void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx) |
d7e28ffe RR |
933 | { |
934 | int pgdir; | |
935 | ||
936 | if (idx >= SWITCHER_PGD_INDEX) | |
937 | return; | |
938 | ||
bff672e6 | 939 | /* If they're talking about a page table we have a shadow for... */ |
ee3db0f2 | 940 | pgdir = find_pgdir(lg, gpgdir); |
d7e28ffe | 941 | if (pgdir < ARRAY_SIZE(lg->pgdirs)) |
bff672e6 | 942 | /* ... throw it away. */ |
90603d15 | 943 | release_pgd(lg->pgdirs[pgdir].pgdir + idx); |
d7e28ffe | 944 | } |
a91d74a3 | 945 | |
acdd0b62 | 946 | #ifdef CONFIG_X86_PAE |
a91d74a3 | 947 | /* For setting a mid-level, we just throw everything away. It's easy. */ |
acdd0b62 MZ |
948 | void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx) |
949 | { | |
950 | guest_pagetable_clear_all(&lg->cpus[0]); | |
951 | } | |
952 | #endif | |
d7e28ffe | 953 | |
2e04ef76 RR |
954 | /*H:500 |
955 | * (vii) Setting up the page tables initially. | |
bff672e6 | 956 | * |
5dea1c88 RR |
957 | * When a Guest is first created, set initialize a shadow page table which |
958 | * we will populate on future faults. The Guest doesn't have any actual | |
959 | * pagetables yet, so we set linear_pages to tell demand_page() to fake it | |
960 | * for the moment. | |
2e04ef76 | 961 | */ |
58a24566 | 962 | int init_guest_pagetable(struct lguest *lg) |
d7e28ffe | 963 | { |
5dea1c88 RR |
964 | struct lg_cpu *cpu = &lg->cpus[0]; |
965 | int allocated = 0; | |
58a24566 | 966 | |
5dea1c88 RR |
967 | /* lg (and lg->cpus[]) starts zeroed: this allocates a new pgdir */ |
968 | cpu->cpu_pgd = new_pgdir(cpu, 0, &allocated); | |
969 | if (!allocated) | |
d7e28ffe | 970 | return -ENOMEM; |
a91d74a3 | 971 | |
5dea1c88 RR |
972 | /* We start with a linear mapping until the initialize. */ |
973 | cpu->linear_pages = true; | |
d7e28ffe RR |
974 | return 0; |
975 | } | |
976 | ||
a91d74a3 | 977 | /*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ |
382ac6b3 | 978 | void page_table_guest_data_init(struct lg_cpu *cpu) |
47436aa4 RR |
979 | { |
980 | /* We get the kernel address: above this is all kernel memory. */ | |
382ac6b3 | 981 | if (get_user(cpu->lg->kernel_address, |
acdd0b62 | 982 | &cpu->lg->lguest_data->kernel_address) |
2e04ef76 RR |
983 | /* |
984 | * We tell the Guest that it can't use the top 2 or 4 MB | |
985 | * of virtual addresses used by the Switcher. | |
986 | */ | |
acdd0b62 | 987 | || put_user(RESERVE_MEM * 1024 * 1024, |
5dea1c88 | 988 | &cpu->lg->lguest_data->reserve_mem)) { |
382ac6b3 | 989 | kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data); |
5dea1c88 RR |
990 | return; |
991 | } | |
47436aa4 | 992 | |
2e04ef76 RR |
993 | /* |
994 | * In flush_user_mappings() we loop from 0 to | |
47436aa4 | 995 | * "pgd_index(lg->kernel_address)". This assumes it won't hit the |
2e04ef76 RR |
996 | * Switcher mappings, so check that now. |
997 | */ | |
68a644d7 | 998 | if (cpu->lg->kernel_address >= switcher_addr) |
382ac6b3 GOC |
999 | kill_guest(cpu, "bad kernel address %#lx", |
1000 | cpu->lg->kernel_address); | |
47436aa4 RR |
1001 | } |
1002 | ||
bff672e6 | 1003 | /* When a Guest dies, our cleanup is fairly simple. */ |
d7e28ffe RR |
1004 | void free_guest_pagetable(struct lguest *lg) |
1005 | { | |
1006 | unsigned int i; | |
1007 | ||
bff672e6 | 1008 | /* Throw away all page table pages. */ |
d7e28ffe | 1009 | release_all_pagetables(lg); |
bff672e6 | 1010 | /* Now free the top levels: free_page() can handle 0 just fine. */ |
d7e28ffe RR |
1011 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
1012 | free_page((long)lg->pgdirs[i].pgdir); | |
1013 | } | |
1014 | ||
2e04ef76 RR |
1015 | /*H:480 |
1016 | * (vi) Mapping the Switcher when the Guest is about to run. | |
bff672e6 | 1017 | * |
e1e72965 | 1018 | * The Switcher and the two pages for this CPU need to be visible in the |
bff672e6 | 1019 | * Guest (and not the pages for other CPUs). We have the appropriate PTE pages |
e1e72965 | 1020 | * for each CPU already set up, we just need to hook them in now we know which |
2e04ef76 RR |
1021 | * Guest is about to run on this CPU. |
1022 | */ | |
0c78441c | 1023 | void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages) |
d7e28ffe | 1024 | { |
c9f29549 | 1025 | pte_t *switcher_pte_page = __this_cpu_read(switcher_pte_pages); |
df29f43e | 1026 | pte_t regs_pte; |
d7e28ffe | 1027 | |
acdd0b62 MZ |
1028 | #ifdef CONFIG_X86_PAE |
1029 | pmd_t switcher_pmd; | |
1030 | pmd_t *pmd_table; | |
1031 | ||
4c1ea3dd RR |
1032 | switcher_pmd = pfn_pmd(__pa(switcher_pte_page) >> PAGE_SHIFT, |
1033 | PAGE_KERNEL_EXEC); | |
acdd0b62 | 1034 | |
a91d74a3 RR |
1035 | /* Figure out where the pmd page is, by reading the PGD, and converting |
1036 | * it to a virtual address. */ | |
acdd0b62 MZ |
1037 | pmd_table = __va(pgd_pfn(cpu->lg-> |
1038 | pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX]) | |
1039 | << PAGE_SHIFT); | |
a91d74a3 | 1040 | /* Now write it into the shadow page table. */ |
4c1ea3dd | 1041 | set_pmd(&pmd_table[SWITCHER_PMD_INDEX], switcher_pmd); |
acdd0b62 MZ |
1042 | #else |
1043 | pgd_t switcher_pgd; | |
1044 | ||
2e04ef76 RR |
1045 | /* |
1046 | * Make the last PGD entry for this Guest point to the Switcher's PTE | |
1047 | * page for this CPU (with appropriate flags). | |
1048 | */ | |
ed1dc778 | 1049 | switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL_EXEC); |
df29f43e | 1050 | |
1713608f | 1051 | cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; |
d7e28ffe | 1052 | |
acdd0b62 | 1053 | #endif |
2e04ef76 RR |
1054 | /* |
1055 | * We also change the Switcher PTE page. When we're running the Guest, | |
bff672e6 RR |
1056 | * we want the Guest's "regs" page to appear where the first Switcher |
1057 | * page for this CPU is. This is an optimization: when the Switcher | |
1058 | * saves the Guest registers, it saves them into the first page of this | |
1059 | * CPU's "struct lguest_pages": if we make sure the Guest's register | |
1060 | * page is already mapped there, we don't have to copy them out | |
2e04ef76 RR |
1061 | * again. |
1062 | */ | |
4c1ea3dd RR |
1063 | regs_pte = pfn_pte(__pa(cpu->regs_page) >> PAGE_SHIFT, PAGE_KERNEL); |
1064 | set_pte(&switcher_pte_page[pte_index((unsigned long)pages)], regs_pte); | |
d7e28ffe | 1065 | } |
bff672e6 | 1066 | /*:*/ |
d7e28ffe RR |
1067 | |
1068 | static void free_switcher_pte_pages(void) | |
1069 | { | |
1070 | unsigned int i; | |
1071 | ||
1072 | for_each_possible_cpu(i) | |
1073 | free_page((long)switcher_pte_page(i)); | |
1074 | } | |
1075 | ||
2e04ef76 RR |
1076 | /*H:520 |
1077 | * Setting up the Switcher PTE page for given CPU is fairly easy, given | |
bff672e6 RR |
1078 | * the CPU number and the "struct page"s for the Switcher code itself. |
1079 | * | |
2e04ef76 RR |
1080 | * Currently the Switcher is less than a page long, so "pages" is always 1. |
1081 | */ | |
d7e28ffe RR |
1082 | static __init void populate_switcher_pte_page(unsigned int cpu, |
1083 | struct page *switcher_page[], | |
1084 | unsigned int pages) | |
1085 | { | |
1086 | unsigned int i; | |
df29f43e | 1087 | pte_t *pte = switcher_pte_page(cpu); |
d7e28ffe | 1088 | |
bff672e6 | 1089 | /* The first entries are easy: they map the Switcher code. */ |
d7e28ffe | 1090 | for (i = 0; i < pages; i++) { |
4c1ea3dd | 1091 | set_pte(&pte[i], mk_pte(switcher_page[i], |
90603d15 | 1092 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED))); |
d7e28ffe RR |
1093 | } |
1094 | ||
bff672e6 | 1095 | /* The only other thing we map is this CPU's pair of pages. */ |
d7e28ffe RR |
1096 | i = pages + cpu*2; |
1097 | ||
bff672e6 | 1098 | /* First page (Guest registers) is writable from the Guest */ |
4c1ea3dd | 1099 | set_pte(&pte[i], pfn_pte(page_to_pfn(switcher_page[i]), |
90603d15 | 1100 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW))); |
df29f43e | 1101 | |
2e04ef76 RR |
1102 | /* |
1103 | * The second page contains the "struct lguest_ro_state", and is | |
1104 | * read-only. | |
1105 | */ | |
4c1ea3dd | 1106 | set_pte(&pte[i+1], pfn_pte(page_to_pfn(switcher_page[i+1]), |
90603d15 | 1107 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED))); |
d7e28ffe RR |
1108 | } |
1109 | ||
2e04ef76 RR |
1110 | /* |
1111 | * We've made it through the page table code. Perhaps our tired brains are | |
e1e72965 RR |
1112 | * still processing the details, or perhaps we're simply glad it's over. |
1113 | * | |
a6bd8e13 RR |
1114 | * If nothing else, note that all this complexity in juggling shadow page tables |
1115 | * in sync with the Guest's page tables is for one reason: for most Guests this | |
1116 | * page table dance determines how bad performance will be. This is why Xen | |
1117 | * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD | |
1118 | * have implemented shadow page table support directly into hardware. | |
e1e72965 | 1119 | * |
2e04ef76 RR |
1120 | * There is just one file remaining in the Host. |
1121 | */ | |
e1e72965 | 1122 | |
2e04ef76 RR |
1123 | /*H:510 |
1124 | * At boot or module load time, init_pagetables() allocates and populates | |
1125 | * the Switcher PTE page for each CPU. | |
1126 | */ | |
d7e28ffe RR |
1127 | __init int init_pagetables(struct page **switcher_page, unsigned int pages) |
1128 | { | |
1129 | unsigned int i; | |
1130 | ||
1131 | for_each_possible_cpu(i) { | |
df29f43e | 1132 | switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL); |
d7e28ffe RR |
1133 | if (!switcher_pte_page(i)) { |
1134 | free_switcher_pte_pages(); | |
1135 | return -ENOMEM; | |
1136 | } | |
1137 | populate_switcher_pte_page(i, switcher_page, pages); | |
1138 | } | |
1139 | return 0; | |
1140 | } | |
bff672e6 | 1141 | /*:*/ |
d7e28ffe | 1142 | |
bff672e6 | 1143 | /* Cleaning up simply involves freeing the PTE page for each CPU. */ |
d7e28ffe RR |
1144 | void free_pagetables(void) |
1145 | { | |
1146 | free_switcher_pte_pages(); | |
1147 | } |