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) | |
acdd0b62 MZ |
66 | #define CHECK_GPGD_MASK _PAGE_PRESENT |
67 | #else | |
acdd0b62 MZ |
68 | #define CHECK_GPGD_MASK _PAGE_TABLE |
69 | #endif | |
70 | ||
2e04ef76 RR |
71 | /* |
72 | * We actually need a separate PTE page for each CPU. Remember that after the | |
bff672e6 | 73 | * Switcher code itself comes two pages for each CPU, and we don't want this |
2e04ef76 RR |
74 | * CPU's guest to see the pages of any other CPU. |
75 | */ | |
df29f43e | 76 | static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); |
d7e28ffe RR |
77 | #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) |
78 | ||
2e04ef76 RR |
79 | /*H:320 |
80 | * The page table code is curly enough to need helper functions to keep it | |
a91d74a3 RR |
81 | * clear and clean. The kernel itself provides many of them; one advantage |
82 | * of insisting that the Guest and Host use the same CONFIG_PAE setting. | |
bff672e6 | 83 | * |
df29f43e | 84 | * There are two functions which return pointers to the shadow (aka "real") |
bff672e6 RR |
85 | * page tables. |
86 | * | |
87 | * spgd_addr() takes the virtual address and returns a pointer to the top-level | |
e1e72965 RR |
88 | * page directory entry (PGD) for that address. Since we keep track of several |
89 | * page tables, the "i" argument tells us which one we're interested in (it's | |
2e04ef76 RR |
90 | * usually the current one). |
91 | */ | |
382ac6b3 | 92 | static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr) |
d7e28ffe | 93 | { |
df29f43e | 94 | unsigned int index = pgd_index(vaddr); |
d7e28ffe | 95 | |
bff672e6 | 96 | /* Return a pointer index'th pgd entry for the i'th page table. */ |
382ac6b3 | 97 | return &cpu->lg->pgdirs[i].pgdir[index]; |
d7e28ffe RR |
98 | } |
99 | ||
acdd0b62 | 100 | #ifdef CONFIG_X86_PAE |
2e04ef76 RR |
101 | /* |
102 | * This routine then takes the PGD entry given above, which contains the | |
acdd0b62 | 103 | * address of the PMD page. It then returns a pointer to the PMD entry for the |
2e04ef76 RR |
104 | * given address. |
105 | */ | |
acdd0b62 MZ |
106 | static pmd_t *spmd_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) |
107 | { | |
108 | unsigned int index = pmd_index(vaddr); | |
109 | pmd_t *page; | |
110 | ||
acdd0b62 MZ |
111 | /* You should never call this if the PGD entry wasn't valid */ |
112 | BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); | |
113 | page = __va(pgd_pfn(spgd) << PAGE_SHIFT); | |
114 | ||
115 | return &page[index]; | |
116 | } | |
117 | #endif | |
118 | ||
2e04ef76 RR |
119 | /* |
120 | * This routine then takes the page directory entry returned above, which | |
e1e72965 | 121 | * contains the address of the page table entry (PTE) page. It then returns a |
2e04ef76 RR |
122 | * pointer to the PTE entry for the given address. |
123 | */ | |
acdd0b62 | 124 | static pte_t *spte_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) |
d7e28ffe | 125 | { |
acdd0b62 MZ |
126 | #ifdef CONFIG_X86_PAE |
127 | pmd_t *pmd = spmd_addr(cpu, spgd, vaddr); | |
128 | pte_t *page = __va(pmd_pfn(*pmd) << PAGE_SHIFT); | |
129 | ||
130 | /* You should never call this if the PMD entry wasn't valid */ | |
131 | BUG_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)); | |
132 | #else | |
df29f43e | 133 | pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT); |
bff672e6 | 134 | /* You should never call this if the PGD entry wasn't valid */ |
df29f43e | 135 | BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); |
acdd0b62 MZ |
136 | #endif |
137 | ||
90603d15 | 138 | return &page[pte_index(vaddr)]; |
d7e28ffe RR |
139 | } |
140 | ||
2e04ef76 | 141 | /* |
9f54288d | 142 | * These functions are just like the above, except they access the Guest |
2e04ef76 RR |
143 | * page tables. Hence they return a Guest address. |
144 | */ | |
1713608f | 145 | static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 146 | { |
df29f43e | 147 | unsigned int index = vaddr >> (PGDIR_SHIFT); |
1713608f | 148 | return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t); |
d7e28ffe RR |
149 | } |
150 | ||
acdd0b62 | 151 | #ifdef CONFIG_X86_PAE |
a91d74a3 | 152 | /* Follow the PGD to the PMD. */ |
acdd0b62 | 153 | static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr) |
d7e28ffe | 154 | { |
df29f43e MZ |
155 | unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; |
156 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); | |
acdd0b62 MZ |
157 | return gpage + pmd_index(vaddr) * sizeof(pmd_t); |
158 | } | |
acdd0b62 | 159 | |
a91d74a3 | 160 | /* Follow the PMD to the PTE. */ |
acdd0b62 | 161 | static unsigned long gpte_addr(struct lg_cpu *cpu, |
92b4d8df | 162 | pmd_t gpmd, unsigned long vaddr) |
acdd0b62 | 163 | { |
92b4d8df | 164 | unsigned long gpage = pmd_pfn(gpmd) << PAGE_SHIFT; |
acdd0b62 | 165 | |
acdd0b62 | 166 | BUG_ON(!(pmd_flags(gpmd) & _PAGE_PRESENT)); |
92b4d8df RR |
167 | return gpage + pte_index(vaddr) * sizeof(pte_t); |
168 | } | |
acdd0b62 | 169 | #else |
a91d74a3 | 170 | /* Follow the PGD to the PTE (no mid-level for !PAE). */ |
92b4d8df RR |
171 | static unsigned long gpte_addr(struct lg_cpu *cpu, |
172 | pgd_t gpgd, unsigned long vaddr) | |
173 | { | |
174 | unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; | |
175 | ||
176 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); | |
90603d15 | 177 | return gpage + pte_index(vaddr) * sizeof(pte_t); |
d7e28ffe | 178 | } |
92b4d8df | 179 | #endif |
a6bd8e13 RR |
180 | /*:*/ |
181 | ||
9f54288d | 182 | /*M:007 |
2e04ef76 RR |
183 | * get_pfn is slow: we could probably try to grab batches of pages here as |
184 | * an optimization (ie. pre-faulting). | |
185 | :*/ | |
d7e28ffe | 186 | |
2e04ef76 RR |
187 | /*H:350 |
188 | * This routine takes a page number given by the Guest and converts it to | |
bff672e6 RR |
189 | * an actual, physical page number. It can fail for several reasons: the |
190 | * virtual address might not be mapped by the Launcher, the write flag is set | |
191 | * and the page is read-only, or the write flag was set and the page was | |
192 | * shared so had to be copied, but we ran out of memory. | |
193 | * | |
a6bd8e13 | 194 | * This holds a reference to the page, so release_pte() is careful to put that |
2e04ef76 RR |
195 | * back. |
196 | */ | |
d7e28ffe RR |
197 | static unsigned long get_pfn(unsigned long virtpfn, int write) |
198 | { | |
199 | struct page *page; | |
71a3f4ed RR |
200 | |
201 | /* gup me one page at this address please! */ | |
202 | if (get_user_pages_fast(virtpfn << PAGE_SHIFT, 1, write, &page) == 1) | |
203 | return page_to_pfn(page); | |
204 | ||
bff672e6 | 205 | /* This value indicates failure. */ |
71a3f4ed | 206 | return -1UL; |
d7e28ffe RR |
207 | } |
208 | ||
2e04ef76 RR |
209 | /*H:340 |
210 | * Converting a Guest page table entry to a shadow (ie. real) page table | |
bff672e6 RR |
211 | * entry can be a little tricky. The flags are (almost) the same, but the |
212 | * Guest PTE contains a virtual page number: the CPU needs the real page | |
2e04ef76 RR |
213 | * number. |
214 | */ | |
382ac6b3 | 215 | static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write) |
d7e28ffe | 216 | { |
df29f43e | 217 | unsigned long pfn, base, flags; |
d7e28ffe | 218 | |
2e04ef76 RR |
219 | /* |
220 | * The Guest sets the global flag, because it thinks that it is using | |
bff672e6 RR |
221 | * PGE. We only told it to use PGE so it would tell us whether it was |
222 | * flushing a kernel mapping or a userspace mapping. We don't actually | |
2e04ef76 RR |
223 | * use the global bit, so throw it away. |
224 | */ | |
df29f43e | 225 | flags = (pte_flags(gpte) & ~_PAGE_GLOBAL); |
bff672e6 | 226 | |
3c6b5bfa | 227 | /* The Guest's pages are offset inside the Launcher. */ |
382ac6b3 | 228 | base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE; |
3c6b5bfa | 229 | |
2e04ef76 RR |
230 | /* |
231 | * We need a temporary "unsigned long" variable to hold the answer from | |
bff672e6 RR |
232 | * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't |
233 | * fit in spte.pfn. get_pfn() finds the real physical number of the | |
2e04ef76 RR |
234 | * page, given the virtual number. |
235 | */ | |
df29f43e | 236 | pfn = get_pfn(base + pte_pfn(gpte), write); |
d7e28ffe | 237 | if (pfn == -1UL) { |
382ac6b3 | 238 | kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte)); |
2e04ef76 RR |
239 | /* |
240 | * When we destroy the Guest, we'll go through the shadow page | |
bff672e6 | 241 | * tables and release_pte() them. Make sure we don't think |
2e04ef76 RR |
242 | * this one is valid! |
243 | */ | |
df29f43e | 244 | flags = 0; |
d7e28ffe | 245 | } |
df29f43e MZ |
246 | /* Now we assemble our shadow PTE from the page number and flags. */ |
247 | return pfn_pte(pfn, __pgprot(flags)); | |
d7e28ffe RR |
248 | } |
249 | ||
bff672e6 | 250 | /*H:460 And to complete the chain, release_pte() looks like this: */ |
df29f43e | 251 | static void release_pte(pte_t pte) |
d7e28ffe | 252 | { |
2e04ef76 RR |
253 | /* |
254 | * Remember that get_user_pages_fast() took a reference to the page, in | |
255 | * get_pfn()? We have to put it back now. | |
256 | */ | |
df29f43e | 257 | if (pte_flags(pte) & _PAGE_PRESENT) |
90603d15 | 258 | put_page(pte_page(pte)); |
d7e28ffe | 259 | } |
bff672e6 | 260 | /*:*/ |
d7e28ffe | 261 | |
e1d12606 | 262 | static bool check_gpte(struct lg_cpu *cpu, pte_t gpte) |
d7e28ffe | 263 | { |
31f4b46e | 264 | if ((pte_flags(gpte) & _PAGE_PSE) || |
e1d12606 | 265 | pte_pfn(gpte) >= cpu->lg->pfn_limit) { |
382ac6b3 | 266 | kill_guest(cpu, "bad page table entry"); |
e1d12606 RR |
267 | return false; |
268 | } | |
269 | return true; | |
d7e28ffe RR |
270 | } |
271 | ||
e1d12606 | 272 | static bool check_gpgd(struct lg_cpu *cpu, pgd_t gpgd) |
d7e28ffe | 273 | { |
acdd0b62 | 274 | if ((pgd_flags(gpgd) & ~CHECK_GPGD_MASK) || |
e1d12606 | 275 | (pgd_pfn(gpgd) >= cpu->lg->pfn_limit)) { |
382ac6b3 | 276 | kill_guest(cpu, "bad page directory entry"); |
e1d12606 RR |
277 | return false; |
278 | } | |
279 | return true; | |
d7e28ffe RR |
280 | } |
281 | ||
acdd0b62 | 282 | #ifdef CONFIG_X86_PAE |
e1d12606 | 283 | static bool check_gpmd(struct lg_cpu *cpu, pmd_t gpmd) |
acdd0b62 MZ |
284 | { |
285 | if ((pmd_flags(gpmd) & ~_PAGE_TABLE) || | |
e1d12606 | 286 | (pmd_pfn(gpmd) >= cpu->lg->pfn_limit)) { |
acdd0b62 | 287 | kill_guest(cpu, "bad page middle directory entry"); |
e1d12606 RR |
288 | return false; |
289 | } | |
290 | return true; | |
acdd0b62 MZ |
291 | } |
292 | #endif | |
293 | ||
17427e08 RR |
294 | /*H:331 |
295 | * This is the core routine to walk the shadow page tables and find the page | |
296 | * table entry for a specific address. | |
297 | * | |
298 | * If allocate is set, then we allocate any missing levels, setting the flags | |
299 | * on the new page directory and mid-level directories using the arguments | |
300 | * (which are copied from the Guest's page table entries). | |
301 | */ | |
302 | static pte_t *find_spte(struct lg_cpu *cpu, unsigned long vaddr, bool allocate, | |
303 | int pgd_flags, int pmd_flags) | |
304 | { | |
305 | pgd_t *spgd; | |
306 | /* Mid level for PAE. */ | |
307 | #ifdef CONFIG_X86_PAE | |
308 | pmd_t *spmd; | |
309 | #endif | |
310 | ||
311 | /* Get top level entry. */ | |
312 | spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); | |
313 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) { | |
314 | /* No shadow entry: allocate a new shadow PTE page. */ | |
315 | unsigned long ptepage; | |
316 | ||
317 | /* If they didn't want us to allocate anything, stop. */ | |
318 | if (!allocate) | |
319 | return NULL; | |
320 | ||
321 | ptepage = get_zeroed_page(GFP_KERNEL); | |
322 | /* | |
323 | * This is not really the Guest's fault, but killing it is | |
324 | * simple for this corner case. | |
325 | */ | |
326 | if (!ptepage) { | |
327 | kill_guest(cpu, "out of memory allocating pte page"); | |
328 | return NULL; | |
329 | } | |
330 | /* | |
331 | * And we copy the flags to the shadow PGD entry. The page | |
332 | * number in the shadow PGD is the page we just allocated. | |
333 | */ | |
334 | set_pgd(spgd, __pgd(__pa(ptepage) | pgd_flags)); | |
335 | } | |
336 | ||
337 | /* | |
338 | * Intel's Physical Address Extension actually uses three levels of | |
339 | * page tables, so we need to look in the mid-level. | |
340 | */ | |
341 | #ifdef CONFIG_X86_PAE | |
342 | /* Now look at the mid-level shadow entry. */ | |
343 | spmd = spmd_addr(cpu, *spgd, vaddr); | |
344 | ||
345 | if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) { | |
346 | /* No shadow entry: allocate a new shadow PTE page. */ | |
347 | unsigned long ptepage; | |
348 | ||
349 | /* If they didn't want us to allocate anything, stop. */ | |
350 | if (!allocate) | |
351 | return NULL; | |
352 | ||
353 | ptepage = get_zeroed_page(GFP_KERNEL); | |
354 | ||
355 | /* | |
356 | * This is not really the Guest's fault, but killing it is | |
357 | * simple for this corner case. | |
358 | */ | |
359 | if (!ptepage) { | |
360 | kill_guest(cpu, "out of memory allocating pmd page"); | |
361 | return NULL; | |
362 | } | |
363 | ||
364 | /* | |
365 | * And we copy the flags to the shadow PMD entry. The page | |
366 | * number in the shadow PMD is the page we just allocated. | |
367 | */ | |
368 | set_pmd(spmd, __pmd(__pa(ptepage) | pmd_flags)); | |
369 | } | |
370 | #endif | |
371 | ||
372 | /* Get the pointer to the shadow PTE entry we're going to set. */ | |
373 | return spte_addr(cpu, *spgd, vaddr); | |
374 | } | |
375 | ||
bff672e6 | 376 | /*H:330 |
e1e72965 | 377 | * (i) Looking up a page table entry when the Guest faults. |
bff672e6 RR |
378 | * |
379 | * We saw this call in run_guest(): when we see a page fault in the Guest, we | |
380 | * come here. That's because we only set up the shadow page tables lazily as | |
381 | * they're needed, so we get page faults all the time and quietly fix them up | |
382 | * and return to the Guest without it knowing. | |
383 | * | |
384 | * If we fixed up the fault (ie. we mapped the address), this routine returns | |
2e04ef76 RR |
385 | * true. Otherwise, it was a real fault and we need to tell the Guest. |
386 | */ | |
df1693ab | 387 | bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) |
d7e28ffe | 388 | { |
d7e28ffe | 389 | unsigned long gpte_ptr; |
df29f43e MZ |
390 | pte_t gpte; |
391 | pte_t *spte; | |
acdd0b62 | 392 | pmd_t gpmd; |
17427e08 | 393 | pgd_t gpgd; |
acdd0b62 | 394 | |
68a644d7 RR |
395 | /* We never demand page the Switcher, so trying is a mistake. */ |
396 | if (vaddr >= switcher_addr) | |
397 | return false; | |
398 | ||
bff672e6 | 399 | /* First step: get the top-level Guest page table entry. */ |
5dea1c88 RR |
400 | if (unlikely(cpu->linear_pages)) { |
401 | /* Faking up a linear mapping. */ | |
402 | gpgd = __pgd(CHECK_GPGD_MASK); | |
403 | } else { | |
404 | gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); | |
405 | /* Toplevel not present? We can't map it in. */ | |
406 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) | |
407 | return false; | |
d7e28ffe | 408 | |
17427e08 RR |
409 | /* |
410 | * This kills the Guest if it has weird flags or tries to | |
411 | * refer to a "physical" address outside the bounds. | |
2e04ef76 | 412 | */ |
e1d12606 RR |
413 | if (!check_gpgd(cpu, gpgd)) |
414 | return false; | |
d7e28ffe RR |
415 | } |
416 | ||
17427e08 RR |
417 | /* This "mid-level" entry is only used for non-linear, PAE mode. */ |
418 | gpmd = __pmd(_PAGE_TABLE); | |
419 | ||
acdd0b62 | 420 | #ifdef CONFIG_X86_PAE |
17427e08 | 421 | if (likely(!cpu->linear_pages)) { |
5dea1c88 RR |
422 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); |
423 | /* Middle level not present? We can't map it in. */ | |
424 | if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) | |
425 | return false; | |
acdd0b62 | 426 | |
17427e08 RR |
427 | /* |
428 | * This kills the Guest if it has weird flags or tries to | |
429 | * refer to a "physical" address outside the bounds. | |
2e04ef76 | 430 | */ |
e1d12606 RR |
431 | if (!check_gpmd(cpu, gpmd)) |
432 | return false; | |
acdd0b62 | 433 | } |
92b4d8df | 434 | |
2e04ef76 RR |
435 | /* |
436 | * OK, now we look at the lower level in the Guest page table: keep its | |
437 | * address, because we might update it later. | |
438 | */ | |
92b4d8df RR |
439 | gpte_ptr = gpte_addr(cpu, gpmd, vaddr); |
440 | #else | |
2e04ef76 RR |
441 | /* |
442 | * OK, now we look at the lower level in the Guest page table: keep its | |
443 | * address, because we might update it later. | |
444 | */ | |
acdd0b62 | 445 | gpte_ptr = gpte_addr(cpu, gpgd, vaddr); |
92b4d8df | 446 | #endif |
a91d74a3 | 447 | |
5dea1c88 RR |
448 | if (unlikely(cpu->linear_pages)) { |
449 | /* Linear? Make up a PTE which points to same page. */ | |
450 | gpte = __pte((vaddr & PAGE_MASK) | _PAGE_RW | _PAGE_PRESENT); | |
451 | } else { | |
452 | /* Read the actual PTE value. */ | |
453 | gpte = lgread(cpu, gpte_ptr, pte_t); | |
454 | } | |
d7e28ffe | 455 | |
bff672e6 | 456 | /* If this page isn't in the Guest page tables, we can't page it in. */ |
df29f43e | 457 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
df1693ab | 458 | return false; |
d7e28ffe | 459 | |
2e04ef76 RR |
460 | /* |
461 | * Check they're not trying to write to a page the Guest wants | |
462 | * read-only (bit 2 of errcode == write). | |
463 | */ | |
df29f43e | 464 | if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW)) |
df1693ab | 465 | return false; |
d7e28ffe | 466 | |
e1e72965 | 467 | /* User access to a kernel-only page? (bit 3 == user access) */ |
df29f43e | 468 | if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER)) |
df1693ab | 469 | return false; |
d7e28ffe | 470 | |
2e04ef76 RR |
471 | /* |
472 | * Check that the Guest PTE flags are OK, and the page number is below | |
473 | * the pfn_limit (ie. not mapping the Launcher binary). | |
474 | */ | |
e1d12606 RR |
475 | if (!check_gpte(cpu, gpte)) |
476 | return false; | |
e1e72965 | 477 | |
bff672e6 | 478 | /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */ |
df29f43e | 479 | gpte = pte_mkyoung(gpte); |
d7e28ffe | 480 | if (errcode & 2) |
df29f43e | 481 | gpte = pte_mkdirty(gpte); |
d7e28ffe | 482 | |
bff672e6 | 483 | /* Get the pointer to the shadow PTE entry we're going to set. */ |
17427e08 RR |
484 | spte = find_spte(cpu, vaddr, true, pgd_flags(gpgd), pmd_flags(gpmd)); |
485 | if (!spte) | |
486 | return false; | |
2e04ef76 RR |
487 | |
488 | /* | |
489 | * If there was a valid shadow PTE entry here before, we release it. | |
490 | * This can happen with a write to a previously read-only entry. | |
491 | */ | |
d7e28ffe RR |
492 | release_pte(*spte); |
493 | ||
2e04ef76 RR |
494 | /* |
495 | * If this is a write, we insist that the Guest page is writable (the | |
496 | * final arg to gpte_to_spte()). | |
497 | */ | |
df29f43e | 498 | if (pte_dirty(gpte)) |
382ac6b3 | 499 | *spte = gpte_to_spte(cpu, gpte, 1); |
df29f43e | 500 | else |
2e04ef76 RR |
501 | /* |
502 | * If this is a read, don't set the "writable" bit in the page | |
bff672e6 | 503 | * table entry, even if the Guest says it's writable. That way |
e1e72965 | 504 | * we will come back here when a write does actually occur, so |
2e04ef76 RR |
505 | * we can update the Guest's _PAGE_DIRTY flag. |
506 | */ | |
4c1ea3dd | 507 | set_pte(spte, gpte_to_spte(cpu, pte_wrprotect(gpte), 0)); |
d7e28ffe | 508 | |
2e04ef76 RR |
509 | /* |
510 | * Finally, we write the Guest PTE entry back: we've set the | |
511 | * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. | |
512 | */ | |
5dea1c88 RR |
513 | if (likely(!cpu->linear_pages)) |
514 | lgwrite(cpu, gpte_ptr, pte_t, gpte); | |
bff672e6 | 515 | |
2e04ef76 RR |
516 | /* |
517 | * The fault is fixed, the page table is populated, the mapping | |
e1e72965 RR |
518 | * manipulated, the result returned and the code complete. A small |
519 | * delay and a trace of alliteration are the only indications the Guest | |
2e04ef76 RR |
520 | * has that a page fault occurred at all. |
521 | */ | |
df1693ab | 522 | return true; |
d7e28ffe RR |
523 | } |
524 | ||
e1e72965 RR |
525 | /*H:360 |
526 | * (ii) Making sure the Guest stack is mapped. | |
bff672e6 | 527 | * |
e1e72965 RR |
528 | * Remember that direct traps into the Guest need a mapped Guest kernel stack. |
529 | * pin_stack_pages() calls us here: we could simply call demand_page(), but as | |
530 | * we've seen that logic is quite long, and usually the stack pages are already | |
531 | * mapped, so it's overkill. | |
bff672e6 RR |
532 | * |
533 | * This is a quick version which answers the question: is this virtual address | |
2e04ef76 RR |
534 | * mapped by the shadow page tables, and is it writable? |
535 | */ | |
df1693ab | 536 | static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 537 | { |
17427e08 | 538 | pte_t *spte; |
d7e28ffe | 539 | unsigned long flags; |
68a644d7 RR |
540 | |
541 | /* You can't put your stack in the Switcher! */ | |
542 | if (vaddr >= switcher_addr) | |
543 | return false; | |
544 | ||
17427e08 RR |
545 | /* If there's no shadow PTE, it's not writable. */ |
546 | spte = find_spte(cpu, vaddr, false, 0, 0); | |
547 | if (!spte) | |
df1693ab | 548 | return false; |
d7e28ffe | 549 | |
2e04ef76 RR |
550 | /* |
551 | * Check the flags on the pte entry itself: it must be present and | |
552 | * writable. | |
553 | */ | |
17427e08 | 554 | flags = pte_flags(*spte); |
d7e28ffe RR |
555 | return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); |
556 | } | |
557 | ||
2e04ef76 RR |
558 | /* |
559 | * So, when pin_stack_pages() asks us to pin a page, we check if it's already | |
bff672e6 | 560 | * in the page tables, and if not, we call demand_page() with error code 2 |
2e04ef76 RR |
561 | * (meaning "write"). |
562 | */ | |
1713608f | 563 | void pin_page(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 564 | { |
1713608f | 565 | if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2)) |
382ac6b3 | 566 | kill_guest(cpu, "bad stack page %#lx", vaddr); |
d7e28ffe | 567 | } |
a91d74a3 | 568 | /*:*/ |
d7e28ffe | 569 | |
acdd0b62 MZ |
570 | #ifdef CONFIG_X86_PAE |
571 | static void release_pmd(pmd_t *spmd) | |
572 | { | |
573 | /* If the entry's not present, there's nothing to release. */ | |
574 | if (pmd_flags(*spmd) & _PAGE_PRESENT) { | |
575 | unsigned int i; | |
576 | pte_t *ptepage = __va(pmd_pfn(*spmd) << PAGE_SHIFT); | |
577 | /* For each entry in the page, we might need to release it. */ | |
578 | for (i = 0; i < PTRS_PER_PTE; i++) | |
579 | release_pte(ptepage[i]); | |
580 | /* Now we can free the page of PTEs */ | |
581 | free_page((long)ptepage); | |
582 | /* And zero out the PMD entry so we never release it twice. */ | |
4c1ea3dd | 583 | set_pmd(spmd, __pmd(0)); |
acdd0b62 MZ |
584 | } |
585 | } | |
586 | ||
587 | static void release_pgd(pgd_t *spgd) | |
588 | { | |
589 | /* If the entry's not present, there's nothing to release. */ | |
590 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { | |
591 | unsigned int i; | |
592 | pmd_t *pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); | |
593 | ||
594 | for (i = 0; i < PTRS_PER_PMD; i++) | |
595 | release_pmd(&pmdpage[i]); | |
596 | ||
597 | /* Now we can free the page of PMDs */ | |
598 | free_page((long)pmdpage); | |
599 | /* And zero out the PGD entry so we never release it twice. */ | |
600 | set_pgd(spgd, __pgd(0)); | |
601 | } | |
602 | } | |
603 | ||
604 | #else /* !CONFIG_X86_PAE */ | |
a91d74a3 RR |
605 | /*H:450 |
606 | * If we chase down the release_pgd() code, the non-PAE version looks like | |
607 | * this. The PAE version is almost identical, but instead of calling | |
608 | * release_pte it calls release_pmd(), which looks much like this. | |
609 | */ | |
90603d15 | 610 | static void release_pgd(pgd_t *spgd) |
d7e28ffe | 611 | { |
bff672e6 | 612 | /* If the entry's not present, there's nothing to release. */ |
df29f43e | 613 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
d7e28ffe | 614 | unsigned int i; |
2e04ef76 RR |
615 | /* |
616 | * Converting the pfn to find the actual PTE page is easy: turn | |
bff672e6 | 617 | * the page number into a physical address, then convert to a |
2e04ef76 RR |
618 | * virtual address (easy for kernel pages like this one). |
619 | */ | |
df29f43e | 620 | pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); |
bff672e6 | 621 | /* For each entry in the page, we might need to release it. */ |
df29f43e | 622 | for (i = 0; i < PTRS_PER_PTE; i++) |
d7e28ffe | 623 | release_pte(ptepage[i]); |
bff672e6 | 624 | /* Now we can free the page of PTEs */ |
d7e28ffe | 625 | free_page((long)ptepage); |
e1e72965 | 626 | /* And zero out the PGD entry so we never release it twice. */ |
df29f43e | 627 | *spgd = __pgd(0); |
d7e28ffe RR |
628 | } |
629 | } | |
acdd0b62 | 630 | #endif |
2e04ef76 RR |
631 | |
632 | /*H:445 | |
633 | * We saw flush_user_mappings() twice: once from the flush_user_mappings() | |
e1e72965 | 634 | * hypercall and once in new_pgdir() when we re-used a top-level pgdir page. |
2e04ef76 RR |
635 | * It simply releases every PTE page from 0 up to the Guest's kernel address. |
636 | */ | |
d7e28ffe RR |
637 | static void flush_user_mappings(struct lguest *lg, int idx) |
638 | { | |
639 | unsigned int i; | |
bff672e6 | 640 | /* Release every pgd entry up to the kernel's address. */ |
47436aa4 | 641 | for (i = 0; i < pgd_index(lg->kernel_address); i++) |
90603d15 | 642 | release_pgd(lg->pgdirs[idx].pgdir + i); |
d7e28ffe RR |
643 | } |
644 | ||
2e04ef76 RR |
645 | /*H:440 |
646 | * (v) Flushing (throwing away) page tables, | |
e1e72965 RR |
647 | * |
648 | * The Guest has a hypercall to throw away the page tables: it's used when a | |
2e04ef76 RR |
649 | * large number of mappings have been changed. |
650 | */ | |
1713608f | 651 | void guest_pagetable_flush_user(struct lg_cpu *cpu) |
d7e28ffe | 652 | { |
bff672e6 | 653 | /* Drop the userspace part of the current page table. */ |
1713608f | 654 | flush_user_mappings(cpu->lg, cpu->cpu_pgd); |
d7e28ffe | 655 | } |
bff672e6 | 656 | /*:*/ |
d7e28ffe | 657 | |
47436aa4 | 658 | /* We walk down the guest page tables to get a guest-physical address */ |
1713608f | 659 | unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr) |
47436aa4 RR |
660 | { |
661 | pgd_t gpgd; | |
662 | pte_t gpte; | |
acdd0b62 MZ |
663 | #ifdef CONFIG_X86_PAE |
664 | pmd_t gpmd; | |
665 | #endif | |
5dea1c88 RR |
666 | |
667 | /* Still not set up? Just map 1:1. */ | |
668 | if (unlikely(cpu->linear_pages)) | |
669 | return vaddr; | |
670 | ||
47436aa4 | 671 | /* First step: get the top-level Guest page table entry. */ |
382ac6b3 | 672 | gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); |
47436aa4 | 673 | /* Toplevel not present? We can't map it in. */ |
6afbdd05 | 674 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) { |
382ac6b3 | 675 | kill_guest(cpu, "Bad address %#lx", vaddr); |
6afbdd05 RR |
676 | return -1UL; |
677 | } | |
47436aa4 | 678 | |
acdd0b62 MZ |
679 | #ifdef CONFIG_X86_PAE |
680 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); | |
681 | if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) | |
682 | kill_guest(cpu, "Bad address %#lx", vaddr); | |
92b4d8df RR |
683 | gpte = lgread(cpu, gpte_addr(cpu, gpmd, vaddr), pte_t); |
684 | #else | |
acdd0b62 | 685 | gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t); |
92b4d8df | 686 | #endif |
47436aa4 | 687 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
382ac6b3 | 688 | kill_guest(cpu, "Bad address %#lx", vaddr); |
47436aa4 RR |
689 | |
690 | return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK); | |
691 | } | |
692 | ||
2e04ef76 RR |
693 | /* |
694 | * We keep several page tables. This is a simple routine to find the page | |
bff672e6 | 695 | * table (if any) corresponding to this top-level address the Guest has given |
2e04ef76 RR |
696 | * us. |
697 | */ | |
d7e28ffe RR |
698 | static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) |
699 | { | |
700 | unsigned int i; | |
701 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) | |
4357bd94 | 702 | if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable) |
d7e28ffe RR |
703 | break; |
704 | return i; | |
705 | } | |
706 | ||
2e04ef76 RR |
707 | /*H:435 |
708 | * And this is us, creating the new page directory. If we really do | |
bff672e6 | 709 | * allocate a new one (and so the kernel parts are not there), we set |
2e04ef76 RR |
710 | * blank_pgdir. |
711 | */ | |
1713608f | 712 | static unsigned int new_pgdir(struct lg_cpu *cpu, |
ee3db0f2 | 713 | unsigned long gpgdir, |
d7e28ffe RR |
714 | int *blank_pgdir) |
715 | { | |
716 | unsigned int next; | |
acdd0b62 MZ |
717 | #ifdef CONFIG_X86_PAE |
718 | pmd_t *pmd_table; | |
719 | #endif | |
d7e28ffe | 720 | |
2e04ef76 RR |
721 | /* |
722 | * We pick one entry at random to throw out. Choosing the Least | |
723 | * Recently Used might be better, but this is easy. | |
724 | */ | |
382ac6b3 | 725 | next = random32() % ARRAY_SIZE(cpu->lg->pgdirs); |
bff672e6 | 726 | /* If it's never been allocated at all before, try now. */ |
382ac6b3 GOC |
727 | if (!cpu->lg->pgdirs[next].pgdir) { |
728 | cpu->lg->pgdirs[next].pgdir = | |
729 | (pgd_t *)get_zeroed_page(GFP_KERNEL); | |
bff672e6 | 730 | /* If the allocation fails, just keep using the one we have */ |
382ac6b3 | 731 | if (!cpu->lg->pgdirs[next].pgdir) |
1713608f | 732 | next = cpu->cpu_pgd; |
acdd0b62 MZ |
733 | else { |
734 | #ifdef CONFIG_X86_PAE | |
2e04ef76 RR |
735 | /* |
736 | * In PAE mode, allocate a pmd page and populate the | |
737 | * last pgd entry. | |
738 | */ | |
acdd0b62 MZ |
739 | pmd_table = (pmd_t *)get_zeroed_page(GFP_KERNEL); |
740 | if (!pmd_table) { | |
741 | free_page((long)cpu->lg->pgdirs[next].pgdir); | |
742 | set_pgd(cpu->lg->pgdirs[next].pgdir, __pgd(0)); | |
743 | next = cpu->cpu_pgd; | |
744 | } else { | |
745 | set_pgd(cpu->lg->pgdirs[next].pgdir + | |
746 | SWITCHER_PGD_INDEX, | |
747 | __pgd(__pa(pmd_table) | _PAGE_PRESENT)); | |
2e04ef76 RR |
748 | /* |
749 | * This is a blank page, so there are no kernel | |
750 | * mappings: caller must map the stack! | |
751 | */ | |
acdd0b62 MZ |
752 | *blank_pgdir = 1; |
753 | } | |
754 | #else | |
d7e28ffe | 755 | *blank_pgdir = 1; |
acdd0b62 MZ |
756 | #endif |
757 | } | |
d7e28ffe | 758 | } |
bff672e6 | 759 | /* Record which Guest toplevel this shadows. */ |
382ac6b3 | 760 | cpu->lg->pgdirs[next].gpgdir = gpgdir; |
d7e28ffe | 761 | /* Release all the non-kernel mappings. */ |
382ac6b3 | 762 | flush_user_mappings(cpu->lg, next); |
d7e28ffe RR |
763 | |
764 | return next; | |
765 | } | |
766 | ||
2e04ef76 RR |
767 | /*H:470 |
768 | * Finally, a routine which throws away everything: all PGD entries in all | |
e1e72965 | 769 | * the shadow page tables, including the Guest's kernel mappings. This is used |
2e04ef76 RR |
770 | * when we destroy the Guest. |
771 | */ | |
d7e28ffe RR |
772 | static void release_all_pagetables(struct lguest *lg) |
773 | { | |
774 | unsigned int i, j; | |
775 | ||
bff672e6 | 776 | /* Every shadow pagetable this Guest has */ |
d7e28ffe | 777 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
acdd0b62 MZ |
778 | if (lg->pgdirs[i].pgdir) { |
779 | #ifdef CONFIG_X86_PAE | |
780 | pgd_t *spgd; | |
781 | pmd_t *pmdpage; | |
782 | unsigned int k; | |
783 | ||
784 | /* Get the last pmd page. */ | |
785 | spgd = lg->pgdirs[i].pgdir + SWITCHER_PGD_INDEX; | |
786 | pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); | |
787 | ||
2e04ef76 RR |
788 | /* |
789 | * And release the pmd entries of that pmd page, | |
790 | * except for the switcher pmd. | |
791 | */ | |
acdd0b62 MZ |
792 | for (k = 0; k < SWITCHER_PMD_INDEX; k++) |
793 | release_pmd(&pmdpage[k]); | |
794 | #endif | |
bff672e6 | 795 | /* Every PGD entry except the Switcher at the top */ |
d7e28ffe | 796 | for (j = 0; j < SWITCHER_PGD_INDEX; j++) |
90603d15 | 797 | release_pgd(lg->pgdirs[i].pgdir + j); |
acdd0b62 | 798 | } |
d7e28ffe RR |
799 | } |
800 | ||
2e04ef76 RR |
801 | /* |
802 | * We also throw away everything when a Guest tells us it's changed a kernel | |
bff672e6 | 803 | * mapping. Since kernel mappings are in every page table, it's easiest to |
e1e72965 | 804 | * throw them all away. This traps the Guest in amber for a while as |
2e04ef76 RR |
805 | * everything faults back in, but it's rare. |
806 | */ | |
4665ac8e | 807 | void guest_pagetable_clear_all(struct lg_cpu *cpu) |
d7e28ffe | 808 | { |
4665ac8e | 809 | release_all_pagetables(cpu->lg); |
bff672e6 | 810 | /* We need the Guest kernel stack mapped again. */ |
4665ac8e | 811 | pin_stack_pages(cpu); |
d7e28ffe | 812 | } |
5dea1c88 RR |
813 | |
814 | /*H:430 | |
815 | * (iv) Switching page tables | |
816 | * | |
817 | * Now we've seen all the page table setting and manipulation, let's see | |
818 | * what happens when the Guest changes page tables (ie. changes the top-level | |
819 | * pgdir). This occurs on almost every context switch. | |
820 | */ | |
821 | void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable) | |
822 | { | |
823 | int newpgdir, repin = 0; | |
824 | ||
825 | /* | |
826 | * The very first time they call this, we're actually running without | |
827 | * any page tables; we've been making it up. Throw them away now. | |
828 | */ | |
829 | if (unlikely(cpu->linear_pages)) { | |
830 | release_all_pagetables(cpu->lg); | |
831 | cpu->linear_pages = false; | |
832 | /* Force allocation of a new pgdir. */ | |
833 | newpgdir = ARRAY_SIZE(cpu->lg->pgdirs); | |
834 | } else { | |
835 | /* Look to see if we have this one already. */ | |
836 | newpgdir = find_pgdir(cpu->lg, pgtable); | |
837 | } | |
838 | ||
839 | /* | |
840 | * If not, we allocate or mug an existing one: if it's a fresh one, | |
841 | * repin gets set to 1. | |
842 | */ | |
843 | if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs)) | |
844 | newpgdir = new_pgdir(cpu, pgtable, &repin); | |
845 | /* Change the current pgd index to the new one. */ | |
846 | cpu->cpu_pgd = newpgdir; | |
847 | /* If it was completely blank, we map in the Guest kernel stack */ | |
848 | if (repin) | |
849 | pin_stack_pages(cpu); | |
850 | } | |
e1e72965 | 851 | /*:*/ |
2e04ef76 RR |
852 | |
853 | /*M:009 | |
854 | * Since we throw away all mappings when a kernel mapping changes, our | |
e1e72965 RR |
855 | * performance sucks for guests using highmem. In fact, a guest with |
856 | * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is | |
857 | * usually slower than a Guest with less memory. | |
858 | * | |
859 | * This, of course, cannot be fixed. It would take some kind of... well, I | |
2e04ef76 RR |
860 | * don't know, but the term "puissant code-fu" comes to mind. |
861 | :*/ | |
d7e28ffe | 862 | |
2e04ef76 RR |
863 | /*H:420 |
864 | * This is the routine which actually sets the page table entry for then | |
bff672e6 RR |
865 | * "idx"'th shadow page table. |
866 | * | |
867 | * Normally, we can just throw out the old entry and replace it with 0: if they | |
868 | * use it demand_page() will put the new entry in. We need to do this anyway: | |
869 | * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page | |
870 | * is read from, and _PAGE_DIRTY when it's written to. | |
871 | * | |
872 | * But Avi Kivity pointed out that most Operating Systems (Linux included) set | |
873 | * these bits on PTEs immediately anyway. This is done to save the CPU from | |
874 | * having to update them, but it helps us the same way: if they set | |
875 | * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if | |
876 | * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately. | |
877 | */ | |
382ac6b3 | 878 | static void do_set_pte(struct lg_cpu *cpu, int idx, |
df29f43e | 879 | unsigned long vaddr, pte_t gpte) |
d7e28ffe | 880 | { |
e1e72965 | 881 | /* Look up the matching shadow page directory entry. */ |
382ac6b3 | 882 | pgd_t *spgd = spgd_addr(cpu, idx, vaddr); |
acdd0b62 MZ |
883 | #ifdef CONFIG_X86_PAE |
884 | pmd_t *spmd; | |
885 | #endif | |
bff672e6 RR |
886 | |
887 | /* If the top level isn't present, there's no entry to update. */ | |
df29f43e | 888 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
acdd0b62 MZ |
889 | #ifdef CONFIG_X86_PAE |
890 | spmd = spmd_addr(cpu, *spgd, vaddr); | |
891 | if (pmd_flags(*spmd) & _PAGE_PRESENT) { | |
892 | #endif | |
2e04ef76 | 893 | /* Otherwise, start by releasing the existing entry. */ |
acdd0b62 MZ |
894 | pte_t *spte = spte_addr(cpu, *spgd, vaddr); |
895 | release_pte(*spte); | |
896 | ||
2e04ef76 RR |
897 | /* |
898 | * If they're setting this entry as dirty or accessed, | |
899 | * we might as well put that entry they've given us in | |
900 | * now. This shaves 10% off a copy-on-write | |
901 | * micro-benchmark. | |
902 | */ | |
acdd0b62 | 903 | if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) { |
e1d12606 RR |
904 | if (!check_gpte(cpu, gpte)) |
905 | return; | |
4c1ea3dd RR |
906 | set_pte(spte, |
907 | gpte_to_spte(cpu, gpte, | |
acdd0b62 | 908 | pte_flags(gpte) & _PAGE_DIRTY)); |
2e04ef76 RR |
909 | } else { |
910 | /* | |
911 | * Otherwise kill it and we can demand_page() | |
912 | * it in later. | |
913 | */ | |
4c1ea3dd | 914 | set_pte(spte, __pte(0)); |
2e04ef76 | 915 | } |
acdd0b62 MZ |
916 | #ifdef CONFIG_X86_PAE |
917 | } | |
918 | #endif | |
d7e28ffe RR |
919 | } |
920 | } | |
921 | ||
2e04ef76 RR |
922 | /*H:410 |
923 | * Updating a PTE entry is a little trickier. | |
bff672e6 RR |
924 | * |
925 | * We keep track of several different page tables (the Guest uses one for each | |
926 | * process, so it makes sense to cache at least a few). Each of these have | |
927 | * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for | |
928 | * all processes. So when the page table above that address changes, we update | |
929 | * all the page tables, not just the current one. This is rare. | |
930 | * | |
a6bd8e13 | 931 | * The benefit is that when we have to track a new page table, we can keep all |
2e04ef76 RR |
932 | * the kernel mappings. This speeds up context switch immensely. |
933 | */ | |
382ac6b3 | 934 | void guest_set_pte(struct lg_cpu *cpu, |
ee3db0f2 | 935 | unsigned long gpgdir, unsigned long vaddr, pte_t gpte) |
d7e28ffe | 936 | { |
68a644d7 RR |
937 | /* We don't let you remap the Switcher; we need it to get back! */ |
938 | if (vaddr >= switcher_addr) { | |
939 | kill_guest(cpu, "attempt to set pte into Switcher pages"); | |
940 | return; | |
941 | } | |
942 | ||
2e04ef76 RR |
943 | /* |
944 | * Kernel mappings must be changed on all top levels. Slow, but doesn't | |
945 | * happen often. | |
946 | */ | |
382ac6b3 | 947 | if (vaddr >= cpu->lg->kernel_address) { |
d7e28ffe | 948 | unsigned int i; |
382ac6b3 GOC |
949 | for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++) |
950 | if (cpu->lg->pgdirs[i].pgdir) | |
951 | do_set_pte(cpu, i, vaddr, gpte); | |
d7e28ffe | 952 | } else { |
bff672e6 | 953 | /* Is this page table one we have a shadow for? */ |
382ac6b3 GOC |
954 | int pgdir = find_pgdir(cpu->lg, gpgdir); |
955 | if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs)) | |
bff672e6 | 956 | /* If so, do the update. */ |
382ac6b3 | 957 | do_set_pte(cpu, pgdir, vaddr, gpte); |
d7e28ffe RR |
958 | } |
959 | } | |
960 | ||
bff672e6 | 961 | /*H:400 |
e1e72965 | 962 | * (iii) Setting up a page table entry when the Guest tells us one has changed. |
bff672e6 RR |
963 | * |
964 | * Just like we did in interrupts_and_traps.c, it makes sense for us to deal | |
965 | * with the other side of page tables while we're here: what happens when the | |
966 | * Guest asks for a page table to be updated? | |
967 | * | |
968 | * We already saw that demand_page() will fill in the shadow page tables when | |
969 | * needed, so we can simply remove shadow page table entries whenever the Guest | |
970 | * tells us they've changed. When the Guest tries to use the new entry it will | |
971 | * fault and demand_page() will fix it up. | |
972 | * | |
fd589a8f | 973 | * So with that in mind here's our code to update a (top-level) PGD entry: |
bff672e6 | 974 | */ |
ebe0ba84 | 975 | void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx) |
d7e28ffe RR |
976 | { |
977 | int pgdir; | |
978 | ||
979 | if (idx >= SWITCHER_PGD_INDEX) | |
980 | return; | |
981 | ||
bff672e6 | 982 | /* If they're talking about a page table we have a shadow for... */ |
ee3db0f2 | 983 | pgdir = find_pgdir(lg, gpgdir); |
d7e28ffe | 984 | if (pgdir < ARRAY_SIZE(lg->pgdirs)) |
bff672e6 | 985 | /* ... throw it away. */ |
90603d15 | 986 | release_pgd(lg->pgdirs[pgdir].pgdir + idx); |
d7e28ffe | 987 | } |
a91d74a3 | 988 | |
acdd0b62 | 989 | #ifdef CONFIG_X86_PAE |
a91d74a3 | 990 | /* For setting a mid-level, we just throw everything away. It's easy. */ |
acdd0b62 MZ |
991 | void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx) |
992 | { | |
993 | guest_pagetable_clear_all(&lg->cpus[0]); | |
994 | } | |
995 | #endif | |
d7e28ffe | 996 | |
2e04ef76 RR |
997 | /*H:500 |
998 | * (vii) Setting up the page tables initially. | |
bff672e6 | 999 | * |
5dea1c88 RR |
1000 | * When a Guest is first created, set initialize a shadow page table which |
1001 | * we will populate on future faults. The Guest doesn't have any actual | |
1002 | * pagetables yet, so we set linear_pages to tell demand_page() to fake it | |
1003 | * for the moment. | |
2e04ef76 | 1004 | */ |
58a24566 | 1005 | int init_guest_pagetable(struct lguest *lg) |
d7e28ffe | 1006 | { |
5dea1c88 RR |
1007 | struct lg_cpu *cpu = &lg->cpus[0]; |
1008 | int allocated = 0; | |
58a24566 | 1009 | |
5dea1c88 RR |
1010 | /* lg (and lg->cpus[]) starts zeroed: this allocates a new pgdir */ |
1011 | cpu->cpu_pgd = new_pgdir(cpu, 0, &allocated); | |
1012 | if (!allocated) | |
d7e28ffe | 1013 | return -ENOMEM; |
a91d74a3 | 1014 | |
5dea1c88 RR |
1015 | /* We start with a linear mapping until the initialize. */ |
1016 | cpu->linear_pages = true; | |
d7e28ffe RR |
1017 | return 0; |
1018 | } | |
1019 | ||
a91d74a3 | 1020 | /*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ |
382ac6b3 | 1021 | void page_table_guest_data_init(struct lg_cpu *cpu) |
47436aa4 | 1022 | { |
c215a8b9 RR |
1023 | /* |
1024 | * We tell the Guest that it can't use the virtual addresses | |
1025 | * used by the Switcher. This trick is equivalent to 4GB - | |
1026 | * switcher_addr. | |
1027 | */ | |
1028 | u32 top = ~switcher_addr + 1; | |
1029 | ||
47436aa4 | 1030 | /* We get the kernel address: above this is all kernel memory. */ |
382ac6b3 | 1031 | if (get_user(cpu->lg->kernel_address, |
c215a8b9 | 1032 | &cpu->lg->lguest_data->kernel_address) |
2e04ef76 | 1033 | /* |
c215a8b9 RR |
1034 | * We tell the Guest that it can't use the top virtual |
1035 | * addresses (used by the Switcher). | |
2e04ef76 | 1036 | */ |
c215a8b9 | 1037 | || put_user(top, &cpu->lg->lguest_data->reserve_mem)) { |
382ac6b3 | 1038 | kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data); |
5dea1c88 RR |
1039 | return; |
1040 | } | |
47436aa4 | 1041 | |
2e04ef76 RR |
1042 | /* |
1043 | * In flush_user_mappings() we loop from 0 to | |
47436aa4 | 1044 | * "pgd_index(lg->kernel_address)". This assumes it won't hit the |
2e04ef76 RR |
1045 | * Switcher mappings, so check that now. |
1046 | */ | |
68a644d7 | 1047 | if (cpu->lg->kernel_address >= switcher_addr) |
382ac6b3 GOC |
1048 | kill_guest(cpu, "bad kernel address %#lx", |
1049 | cpu->lg->kernel_address); | |
47436aa4 RR |
1050 | } |
1051 | ||
bff672e6 | 1052 | /* When a Guest dies, our cleanup is fairly simple. */ |
d7e28ffe RR |
1053 | void free_guest_pagetable(struct lguest *lg) |
1054 | { | |
1055 | unsigned int i; | |
1056 | ||
bff672e6 | 1057 | /* Throw away all page table pages. */ |
d7e28ffe | 1058 | release_all_pagetables(lg); |
bff672e6 | 1059 | /* Now free the top levels: free_page() can handle 0 just fine. */ |
d7e28ffe RR |
1060 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
1061 | free_page((long)lg->pgdirs[i].pgdir); | |
1062 | } | |
1063 | ||
2e04ef76 RR |
1064 | /*H:480 |
1065 | * (vi) Mapping the Switcher when the Guest is about to run. | |
bff672e6 | 1066 | * |
e1e72965 | 1067 | * The Switcher and the two pages for this CPU need to be visible in the |
bff672e6 | 1068 | * Guest (and not the pages for other CPUs). We have the appropriate PTE pages |
e1e72965 | 1069 | * for each CPU already set up, we just need to hook them in now we know which |
2e04ef76 RR |
1070 | * Guest is about to run on this CPU. |
1071 | */ | |
0c78441c | 1072 | void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages) |
d7e28ffe | 1073 | { |
c9f29549 | 1074 | pte_t *switcher_pte_page = __this_cpu_read(switcher_pte_pages); |
df29f43e | 1075 | pte_t regs_pte; |
d7e28ffe | 1076 | |
acdd0b62 MZ |
1077 | #ifdef CONFIG_X86_PAE |
1078 | pmd_t switcher_pmd; | |
1079 | pmd_t *pmd_table; | |
1080 | ||
4c1ea3dd RR |
1081 | switcher_pmd = pfn_pmd(__pa(switcher_pte_page) >> PAGE_SHIFT, |
1082 | PAGE_KERNEL_EXEC); | |
acdd0b62 | 1083 | |
a91d74a3 RR |
1084 | /* Figure out where the pmd page is, by reading the PGD, and converting |
1085 | * it to a virtual address. */ | |
acdd0b62 MZ |
1086 | pmd_table = __va(pgd_pfn(cpu->lg-> |
1087 | pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX]) | |
1088 | << PAGE_SHIFT); | |
a91d74a3 | 1089 | /* Now write it into the shadow page table. */ |
4c1ea3dd | 1090 | set_pmd(&pmd_table[SWITCHER_PMD_INDEX], switcher_pmd); |
acdd0b62 MZ |
1091 | #else |
1092 | pgd_t switcher_pgd; | |
1093 | ||
2e04ef76 RR |
1094 | /* |
1095 | * Make the last PGD entry for this Guest point to the Switcher's PTE | |
1096 | * page for this CPU (with appropriate flags). | |
1097 | */ | |
ed1dc778 | 1098 | switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL_EXEC); |
df29f43e | 1099 | |
1713608f | 1100 | cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; |
d7e28ffe | 1101 | |
acdd0b62 | 1102 | #endif |
2e04ef76 RR |
1103 | /* |
1104 | * We also change the Switcher PTE page. When we're running the Guest, | |
bff672e6 RR |
1105 | * we want the Guest's "regs" page to appear where the first Switcher |
1106 | * page for this CPU is. This is an optimization: when the Switcher | |
1107 | * saves the Guest registers, it saves them into the first page of this | |
1108 | * CPU's "struct lguest_pages": if we make sure the Guest's register | |
1109 | * page is already mapped there, we don't have to copy them out | |
2e04ef76 RR |
1110 | * again. |
1111 | */ | |
4c1ea3dd RR |
1112 | regs_pte = pfn_pte(__pa(cpu->regs_page) >> PAGE_SHIFT, PAGE_KERNEL); |
1113 | set_pte(&switcher_pte_page[pte_index((unsigned long)pages)], regs_pte); | |
d7e28ffe | 1114 | } |
bff672e6 | 1115 | /*:*/ |
d7e28ffe RR |
1116 | |
1117 | static void free_switcher_pte_pages(void) | |
1118 | { | |
1119 | unsigned int i; | |
1120 | ||
1121 | for_each_possible_cpu(i) | |
1122 | free_page((long)switcher_pte_page(i)); | |
1123 | } | |
1124 | ||
2e04ef76 RR |
1125 | /*H:520 |
1126 | * Setting up the Switcher PTE page for given CPU is fairly easy, given | |
93a2cdff | 1127 | * the CPU number and the "struct page"s for the Switcher and per-cpu pages. |
2e04ef76 | 1128 | */ |
d7e28ffe | 1129 | static __init void populate_switcher_pte_page(unsigned int cpu, |
93a2cdff | 1130 | struct page *switcher_pages[]) |
d7e28ffe | 1131 | { |
df29f43e | 1132 | pte_t *pte = switcher_pte_page(cpu); |
93a2cdff | 1133 | int i; |
d7e28ffe | 1134 | |
93a2cdff RR |
1135 | /* The first entries maps the Switcher code. */ |
1136 | set_pte(&pte[0], mk_pte(switcher_pages[0], | |
90603d15 | 1137 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED))); |
d7e28ffe | 1138 | |
bff672e6 | 1139 | /* The only other thing we map is this CPU's pair of pages. */ |
93a2cdff | 1140 | i = 1 + cpu*2; |
d7e28ffe | 1141 | |
bff672e6 | 1142 | /* First page (Guest registers) is writable from the Guest */ |
856c6088 | 1143 | set_pte(&pte[i], pfn_pte(page_to_pfn(switcher_pages[i]), |
90603d15 | 1144 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW))); |
df29f43e | 1145 | |
2e04ef76 RR |
1146 | /* |
1147 | * The second page contains the "struct lguest_ro_state", and is | |
1148 | * read-only. | |
1149 | */ | |
856c6088 | 1150 | set_pte(&pte[i+1], pfn_pte(page_to_pfn(switcher_pages[i+1]), |
90603d15 | 1151 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED))); |
d7e28ffe RR |
1152 | } |
1153 | ||
2e04ef76 RR |
1154 | /* |
1155 | * We've made it through the page table code. Perhaps our tired brains are | |
e1e72965 RR |
1156 | * still processing the details, or perhaps we're simply glad it's over. |
1157 | * | |
a6bd8e13 RR |
1158 | * If nothing else, note that all this complexity in juggling shadow page tables |
1159 | * in sync with the Guest's page tables is for one reason: for most Guests this | |
1160 | * page table dance determines how bad performance will be. This is why Xen | |
1161 | * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD | |
1162 | * have implemented shadow page table support directly into hardware. | |
e1e72965 | 1163 | * |
2e04ef76 RR |
1164 | * There is just one file remaining in the Host. |
1165 | */ | |
e1e72965 | 1166 | |
2e04ef76 RR |
1167 | /*H:510 |
1168 | * At boot or module load time, init_pagetables() allocates and populates | |
1169 | * the Switcher PTE page for each CPU. | |
1170 | */ | |
93a2cdff | 1171 | __init int init_pagetables(struct page **switcher_pages) |
d7e28ffe RR |
1172 | { |
1173 | unsigned int i; | |
1174 | ||
1175 | for_each_possible_cpu(i) { | |
df29f43e | 1176 | switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL); |
d7e28ffe RR |
1177 | if (!switcher_pte_page(i)) { |
1178 | free_switcher_pte_pages(); | |
1179 | return -ENOMEM; | |
1180 | } | |
93a2cdff | 1181 | populate_switcher_pte_page(i, switcher_pages); |
d7e28ffe RR |
1182 | } |
1183 | return 0; | |
1184 | } | |
bff672e6 | 1185 | /*:*/ |
d7e28ffe | 1186 | |
bff672e6 | 1187 | /* Cleaning up simply involves freeing the PTE page for each CPU. */ |
d7e28ffe RR |
1188 | void free_pagetables(void) |
1189 | { | |
1190 | free_switcher_pte_pages(); | |
1191 | } |