aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/lguest/page_tables.c
diff options
context:
space:
mode:
authorRusty Russell <rusty@rustcorp.com.au>2007-07-26 13:41:04 -0400
committerLinus Torvalds <torvalds@woody.linux-foundation.org>2007-07-26 14:35:17 -0400
commitbff672e630a015d5b54c8bfb16160b7edc39a57c (patch)
tree3af06baacb76809234a3e71033d14b7ed769dbd8 /drivers/lguest/page_tables.c
parentdde797899ac17ebb812b7566044124d785e98dc7 (diff)
lguest: documentation V: Host
Documentation: The Host Signed-off-by: Rusty Russell <rusty@rustcorp.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'drivers/lguest/page_tables.c')
-rw-r--r--drivers/lguest/page_tables.c314
1 files changed, 286 insertions, 28 deletions
diff --git a/drivers/lguest/page_tables.c b/drivers/lguest/page_tables.c
index f9ca50d8046..cd047e81cd6 100644
--- a/drivers/lguest/page_tables.c
+++ b/drivers/lguest/page_tables.c
@@ -15,38 +15,91 @@
15#include <asm/tlbflush.h> 15#include <asm/tlbflush.h>
16#include "lg.h" 16#include "lg.h"
17 17
18/*H:300
19 * The Page Table Code
20 *
21 * We use two-level page tables for the Guest. If you're not entirely
22 * comfortable with virtual addresses, physical addresses and page tables then
23 * I recommend you review lguest.c's "Page Table Handling" (with diagrams!).
24 *
25 * The Guest keeps page tables, but we maintain the actual ones here: these are
26 * called "shadow" page tables. Which is a very Guest-centric name: these are
27 * the real page tables the CPU uses, although we keep them up to date to
28 * reflect the Guest's. (See what I mean about weird naming? Since when do
29 * shadows reflect anything?)
30 *
31 * Anyway, this is the most complicated part of the Host code. There are seven
32 * parts to this:
33 * (i) Setting up a page table entry for the Guest when it faults,
34 * (ii) Setting up the page table entry for the Guest stack,
35 * (iii) Setting up a page table entry when the Guest tells us it has changed,
36 * (iv) Switching page tables,
37 * (v) Flushing (thowing away) page tables,
38 * (vi) Mapping the Switcher when the Guest is about to run,
39 * (vii) Setting up the page tables initially.
40 :*/
41
42/* Pages a 4k long, and each page table entry is 4 bytes long, giving us 1024
43 * (or 2^10) entries per page. */
18#define PTES_PER_PAGE_SHIFT 10 44#define PTES_PER_PAGE_SHIFT 10
19#define PTES_PER_PAGE (1 << PTES_PER_PAGE_SHIFT) 45#define PTES_PER_PAGE (1 << PTES_PER_PAGE_SHIFT)
46
47/* 1024 entries in a page table page maps 1024 pages: 4MB. The Switcher is
48 * conveniently placed at the top 4MB, so it uses a separate, complete PTE
49 * page. */
20#define SWITCHER_PGD_INDEX (PTES_PER_PAGE - 1) 50#define SWITCHER_PGD_INDEX (PTES_PER_PAGE - 1)
21 51
52/* We actually need a separate PTE page for each CPU. Remember that after the
53 * Switcher code itself comes two pages for each CPU, and we don't want this
54 * CPU's guest to see the pages of any other CPU. */
22static DEFINE_PER_CPU(spte_t *, switcher_pte_pages); 55static DEFINE_PER_CPU(spte_t *, switcher_pte_pages);
23#define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) 56#define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
24 57
58/*H:320 With our shadow and Guest types established, we need to deal with
59 * them: the page table code is curly enough to need helper functions to keep
60 * it clear and clean.
61 *
62 * The first helper takes a virtual address, and says which entry in the top
63 * level page table deals with that address. Since each top level entry deals
64 * with 4M, this effectively divides by 4M. */
25static unsigned vaddr_to_pgd_index(unsigned long vaddr) 65static unsigned vaddr_to_pgd_index(unsigned long vaddr)
26{ 66{
27 return vaddr >> (PAGE_SHIFT + PTES_PER_PAGE_SHIFT); 67 return vaddr >> (PAGE_SHIFT + PTES_PER_PAGE_SHIFT);
28} 68}
29 69
30/* These access the shadow versions (ie. the ones used by the CPU). */ 70/* There are two functions which return pointers to the shadow (aka "real")
71 * page tables.
72 *
73 * spgd_addr() takes the virtual address and returns a pointer to the top-level
74 * page directory entry for that address. Since we keep track of several page
75 * tables, the "i" argument tells us which one we're interested in (it's
76 * usually the current one). */
31static spgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr) 77static spgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr)
32{ 78{
33 unsigned int index = vaddr_to_pgd_index(vaddr); 79 unsigned int index = vaddr_to_pgd_index(vaddr);
34 80
81 /* We kill any Guest trying to touch the Switcher addresses. */
35 if (index >= SWITCHER_PGD_INDEX) { 82 if (index >= SWITCHER_PGD_INDEX) {
36 kill_guest(lg, "attempt to access switcher pages"); 83 kill_guest(lg, "attempt to access switcher pages");
37 index = 0; 84 index = 0;
38 } 85 }
86 /* Return a pointer index'th pgd entry for the i'th page table. */
39 return &lg->pgdirs[i].pgdir[index]; 87 return &lg->pgdirs[i].pgdir[index];
40} 88}
41 89
90/* This routine then takes the PGD entry given above, which contains the
91 * address of the PTE page. It then returns a pointer to the PTE entry for the
92 * given address. */
42static spte_t *spte_addr(struct lguest *lg, spgd_t spgd, unsigned long vaddr) 93static spte_t *spte_addr(struct lguest *lg, spgd_t spgd, unsigned long vaddr)
43{ 94{
44 spte_t *page = __va(spgd.pfn << PAGE_SHIFT); 95 spte_t *page = __va(spgd.pfn << PAGE_SHIFT);
96 /* You should never call this if the PGD entry wasn't valid */
45 BUG_ON(!(spgd.flags & _PAGE_PRESENT)); 97 BUG_ON(!(spgd.flags & _PAGE_PRESENT));
46 return &page[(vaddr >> PAGE_SHIFT) % PTES_PER_PAGE]; 98 return &page[(vaddr >> PAGE_SHIFT) % PTES_PER_PAGE];
47} 99}
48 100
49/* These access the guest versions. */ 101/* These two functions just like the above two, except they access the Guest
102 * page tables. Hence they return a Guest address. */
50static unsigned long gpgd_addr(struct lguest *lg, unsigned long vaddr) 103static unsigned long gpgd_addr(struct lguest *lg, unsigned long vaddr)
51{ 104{
52 unsigned int index = vaddr >> (PAGE_SHIFT + PTES_PER_PAGE_SHIFT); 105 unsigned int index = vaddr >> (PAGE_SHIFT + PTES_PER_PAGE_SHIFT);
@@ -61,12 +114,24 @@ static unsigned long gpte_addr(struct lguest *lg,
61 return gpage + ((vaddr>>PAGE_SHIFT) % PTES_PER_PAGE) * sizeof(gpte_t); 114 return gpage + ((vaddr>>PAGE_SHIFT) % PTES_PER_PAGE) * sizeof(gpte_t);
62} 115}
63 116
64/* Do a virtual -> physical mapping on a user page. */ 117/*H:350 This routine takes a page number given by the Guest and converts it to
118 * an actual, physical page number. It can fail for several reasons: the
119 * virtual address might not be mapped by the Launcher, the write flag is set
120 * and the page is read-only, or the write flag was set and the page was
121 * shared so had to be copied, but we ran out of memory.
122 *
123 * This holds a reference to the page, so release_pte() is careful to
124 * put that back. */
65static unsigned long get_pfn(unsigned long virtpfn, int write) 125static unsigned long get_pfn(unsigned long virtpfn, int write)
66{ 126{
67 struct page *page; 127 struct page *page;
128 /* This value indicates failure. */
68 unsigned long ret = -1UL; 129 unsigned long ret = -1UL;
69 130
131 /* get_user_pages() is a complex interface: it gets the "struct
132 * vm_area_struct" and "struct page" assocated with a range of pages.
133 * It also needs the task's mmap_sem held, and is not very quick.
134 * It returns the number of pages it got. */
70 down_read(&current->mm->mmap_sem); 135 down_read(&current->mm->mmap_sem);
71 if (get_user_pages(current, current->mm, virtpfn << PAGE_SHIFT, 136 if (get_user_pages(current, current->mm, virtpfn << PAGE_SHIFT,
72 1, write, 1, &page, NULL) == 1) 137 1, write, 1, &page, NULL) == 1)
@@ -75,28 +140,47 @@ static unsigned long get_pfn(unsigned long virtpfn, int write)
75 return ret; 140 return ret;
76} 141}
77 142
143/*H:340 Converting a Guest page table entry to a shadow (ie. real) page table
144 * entry can be a little tricky. The flags are (almost) the same, but the
145 * Guest PTE contains a virtual page number: the CPU needs the real page
146 * number. */
78static spte_t gpte_to_spte(struct lguest *lg, gpte_t gpte, int write) 147static spte_t gpte_to_spte(struct lguest *lg, gpte_t gpte, int write)
79{ 148{
80 spte_t spte; 149 spte_t spte;
81 unsigned long pfn; 150 unsigned long pfn;
82 151
83 /* We ignore the global flag. */ 152 /* The Guest sets the global flag, because it thinks that it is using
153 * PGE. We only told it to use PGE so it would tell us whether it was
154 * flushing a kernel mapping or a userspace mapping. We don't actually
155 * use the global bit, so throw it away. */
84 spte.flags = (gpte.flags & ~_PAGE_GLOBAL); 156 spte.flags = (gpte.flags & ~_PAGE_GLOBAL);
157
158 /* We need a temporary "unsigned long" variable to hold the answer from
159 * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
160 * fit in spte.pfn. get_pfn() finds the real physical number of the
161 * page, given the virtual number. */
85 pfn = get_pfn(gpte.pfn, write); 162 pfn = get_pfn(gpte.pfn, write);
86 if (pfn == -1UL) { 163 if (pfn == -1UL) {
87 kill_guest(lg, "failed to get page %u", gpte.pfn); 164 kill_guest(lg, "failed to get page %u", gpte.pfn);
88 /* Must not put_page() bogus page on cleanup. */ 165 /* When we destroy the Guest, we'll go through the shadow page
166 * tables and release_pte() them. Make sure we don't think
167 * this one is valid! */
89 spte.flags = 0; 168 spte.flags = 0;
90 } 169 }
170 /* Now we assign the page number, and our shadow PTE is complete. */
91 spte.pfn = pfn; 171 spte.pfn = pfn;
92 return spte; 172 return spte;
93} 173}
94 174
175/*H:460 And to complete the chain, release_pte() looks like this: */
95static void release_pte(spte_t pte) 176static void release_pte(spte_t pte)
96{ 177{
178 /* Remember that get_user_pages() took a reference to the page, in
179 * get_pfn()? We have to put it back now. */
97 if (pte.flags & _PAGE_PRESENT) 180 if (pte.flags & _PAGE_PRESENT)
98 put_page(pfn_to_page(pte.pfn)); 181 put_page(pfn_to_page(pte.pfn));
99} 182}
183/*:*/
100 184
101static void check_gpte(struct lguest *lg, gpte_t gpte) 185static void check_gpte(struct lguest *lg, gpte_t gpte)
102{ 186{
@@ -110,11 +194,16 @@ static void check_gpgd(struct lguest *lg, gpgd_t gpgd)
110 kill_guest(lg, "bad page directory entry"); 194 kill_guest(lg, "bad page directory entry");
111} 195}
112 196
113/* FIXME: We hold reference to pages, which prevents them from being 197/*H:330
114 swapped. It'd be nice to have a callback when Linux wants to swap out. */ 198 * (i) Setting up a page table entry for the Guest when it faults
115 199 *
116/* We fault pages in, which allows us to update accessed/dirty bits. 200 * We saw this call in run_guest(): when we see a page fault in the Guest, we
117 * Return true if we got page. */ 201 * come here. That's because we only set up the shadow page tables lazily as
202 * they're needed, so we get page faults all the time and quietly fix them up
203 * and return to the Guest without it knowing.
204 *
205 * If we fixed up the fault (ie. we mapped the address), this routine returns
206 * true. */
118int demand_page(struct lguest *lg, unsigned long vaddr, int errcode) 207int demand_page(struct lguest *lg, unsigned long vaddr, int errcode)
119{ 208{
120 gpgd_t gpgd; 209 gpgd_t gpgd;
@@ -123,106 +212,161 @@ int demand_page(struct lguest *lg, unsigned long vaddr, int errcode)
123 gpte_t gpte; 212 gpte_t gpte;
124 spte_t *spte; 213 spte_t *spte;
125 214
215 /* First step: get the top-level Guest page table entry. */
126 gpgd = mkgpgd(lgread_u32(lg, gpgd_addr(lg, vaddr))); 216 gpgd = mkgpgd(lgread_u32(lg, gpgd_addr(lg, vaddr)));
217 /* Toplevel not present? We can't map it in. */
127 if (!(gpgd.flags & _PAGE_PRESENT)) 218 if (!(gpgd.flags & _PAGE_PRESENT))
128 return 0; 219 return 0;
129 220
221 /* Now look at the matching shadow entry. */
130 spgd = spgd_addr(lg, lg->pgdidx, vaddr); 222 spgd = spgd_addr(lg, lg->pgdidx, vaddr);
131 if (!(spgd->flags & _PAGE_PRESENT)) { 223 if (!(spgd->flags & _PAGE_PRESENT)) {
132 /* Get a page of PTEs for them. */ 224 /* No shadow entry: allocate a new shadow PTE page. */
133 unsigned long ptepage = get_zeroed_page(GFP_KERNEL); 225 unsigned long ptepage = get_zeroed_page(GFP_KERNEL);
134 /* FIXME: Steal from self in this case? */ 226 /* This is not really the Guest's fault, but killing it is
227 * simple for this corner case. */
135 if (!ptepage) { 228 if (!ptepage) {
136 kill_guest(lg, "out of memory allocating pte page"); 229 kill_guest(lg, "out of memory allocating pte page");
137 return 0; 230 return 0;
138 } 231 }
232 /* We check that the Guest pgd is OK. */
139 check_gpgd(lg, gpgd); 233 check_gpgd(lg, gpgd);
234 /* And we copy the flags to the shadow PGD entry. The page
235 * number in the shadow PGD is the page we just allocated. */
140 spgd->raw.val = (__pa(ptepage) | gpgd.flags); 236 spgd->raw.val = (__pa(ptepage) | gpgd.flags);
141 } 237 }
142 238
239 /* OK, now we look at the lower level in the Guest page table: keep its
240 * address, because we might update it later. */
143 gpte_ptr = gpte_addr(lg, gpgd, vaddr); 241 gpte_ptr = gpte_addr(lg, gpgd, vaddr);
144 gpte = mkgpte(lgread_u32(lg, gpte_ptr)); 242 gpte = mkgpte(lgread_u32(lg, gpte_ptr));
145 243
146 /* No page? */ 244 /* If this page isn't in the Guest page tables, we can't page it in. */
147 if (!(gpte.flags & _PAGE_PRESENT)) 245 if (!(gpte.flags & _PAGE_PRESENT))
148 return 0; 246 return 0;
149 247
150 /* Write to read-only page? */ 248 /* Check they're not trying to write to a page the Guest wants
249 * read-only (bit 2 of errcode == write). */
151 if ((errcode & 2) && !(gpte.flags & _PAGE_RW)) 250 if ((errcode & 2) && !(gpte.flags & _PAGE_RW))
152 return 0; 251 return 0;
153 252
154 /* User access to a non-user page? */ 253 /* User access to a kernel page? (bit 3 == user access) */
155 if ((errcode & 4) && !(gpte.flags & _PAGE_USER)) 254 if ((errcode & 4) && !(gpte.flags & _PAGE_USER))
156 return 0; 255 return 0;
157 256
257 /* Check that the Guest PTE flags are OK, and the page number is below
258 * the pfn_limit (ie. not mapping the Launcher binary). */
158 check_gpte(lg, gpte); 259 check_gpte(lg, gpte);
260 /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
159 gpte.flags |= _PAGE_ACCESSED; 261 gpte.flags |= _PAGE_ACCESSED;
160 if (errcode & 2) 262 if (errcode & 2)
161 gpte.flags |= _PAGE_DIRTY; 263 gpte.flags |= _PAGE_DIRTY;
162 264
163 /* We're done with the old pte. */ 265 /* Get the pointer to the shadow PTE entry we're going to set. */
164 spte = spte_addr(lg, *spgd, vaddr); 266 spte = spte_addr(lg, *spgd, vaddr);
267 /* If there was a valid shadow PTE entry here before, we release it.
268 * This can happen with a write to a previously read-only entry. */
165 release_pte(*spte); 269 release_pte(*spte);
166 270
167 /* We don't make it writable if this isn't a write: later 271 /* If this is a write, we insist that the Guest page is writable (the
168 * write will fault so we can set dirty bit in guest. */ 272 * final arg to gpte_to_spte()). */
169 if (gpte.flags & _PAGE_DIRTY) 273 if (gpte.flags & _PAGE_DIRTY)
170 *spte = gpte_to_spte(lg, gpte, 1); 274 *spte = gpte_to_spte(lg, gpte, 1);
171 else { 275 else {
276 /* If this is a read, don't set the "writable" bit in the page
277 * table entry, even if the Guest says it's writable. That way
278 * we come back here when a write does actually ocur, so we can
279 * update the Guest's _PAGE_DIRTY flag. */
172 gpte_t ro_gpte = gpte; 280 gpte_t ro_gpte = gpte;
173 ro_gpte.flags &= ~_PAGE_RW; 281 ro_gpte.flags &= ~_PAGE_RW;
174 *spte = gpte_to_spte(lg, ro_gpte, 0); 282 *spte = gpte_to_spte(lg, ro_gpte, 0);
175 } 283 }
176 284
177 /* Now we update dirty/accessed on guest. */ 285 /* Finally, we write the Guest PTE entry back: we've set the
286 * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */
178 lgwrite_u32(lg, gpte_ptr, gpte.raw.val); 287 lgwrite_u32(lg, gpte_ptr, gpte.raw.val);
288
289 /* We succeeded in mapping the page! */
179 return 1; 290 return 1;
180} 291}
181 292
182/* This is much faster than the full demand_page logic. */ 293/*H:360 (ii) Setting up the page table entry for the Guest stack.
294 *
295 * Remember pin_stack_pages() which makes sure the stack is mapped? It could
296 * simply call demand_page(), but as we've seen that logic is quite long, and
297 * usually the stack pages are already mapped anyway, so it's not required.
298 *
299 * This is a quick version which answers the question: is this virtual address
300 * mapped by the shadow page tables, and is it writable? */
183static int page_writable(struct lguest *lg, unsigned long vaddr) 301static int page_writable(struct lguest *lg, unsigned long vaddr)
184{ 302{
185 spgd_t *spgd; 303 spgd_t *spgd;
186 unsigned long flags; 304 unsigned long flags;
187 305
306 /* Look at the top level entry: is it present? */
188 spgd = spgd_addr(lg, lg->pgdidx, vaddr); 307 spgd = spgd_addr(lg, lg->pgdidx, vaddr);
189 if (!(spgd->flags & _PAGE_PRESENT)) 308 if (!(spgd->flags & _PAGE_PRESENT))
190 return 0; 309 return 0;
191 310
311 /* Check the flags on the pte entry itself: it must be present and
312 * writable. */
192 flags = spte_addr(lg, *spgd, vaddr)->flags; 313 flags = spte_addr(lg, *spgd, vaddr)->flags;
193 return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); 314 return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW);
194} 315}
195 316
317/* So, when pin_stack_pages() asks us to pin a page, we check if it's already
318 * in the page tables, and if not, we call demand_page() with error code 2
319 * (meaning "write"). */
196void pin_page(struct lguest *lg, unsigned long vaddr) 320void pin_page(struct lguest *lg, unsigned long vaddr)
197{ 321{
198 if (!page_writable(lg, vaddr) && !demand_page(lg, vaddr, 2)) 322 if (!page_writable(lg, vaddr) && !demand_page(lg, vaddr, 2))
199 kill_guest(lg, "bad stack page %#lx", vaddr); 323 kill_guest(lg, "bad stack page %#lx", vaddr);
200} 324}
201 325
326/*H:450 If we chase down the release_pgd() code, it looks like this: */
202static void release_pgd(struct lguest *lg, spgd_t *spgd) 327static void release_pgd(struct lguest *lg, spgd_t *spgd)
203{ 328{
329 /* If the entry's not present, there's nothing to release. */
204 if (spgd->flags & _PAGE_PRESENT) { 330 if (spgd->flags & _PAGE_PRESENT) {
205 unsigned int i; 331 unsigned int i;
332 /* Converting the pfn to find the actual PTE page is easy: turn
333 * the page number into a physical address, then convert to a
334 * virtual address (easy for kernel pages like this one). */
206 spte_t *ptepage = __va(spgd->pfn << PAGE_SHIFT); 335 spte_t *ptepage = __va(spgd->pfn << PAGE_SHIFT);
336 /* For each entry in the page, we might need to release it. */
207 for (i = 0; i < PTES_PER_PAGE; i++) 337 for (i = 0; i < PTES_PER_PAGE; i++)
208 release_pte(ptepage[i]); 338 release_pte(ptepage[i]);
339 /* Now we can free the page of PTEs */
209 free_page((long)ptepage); 340 free_page((long)ptepage);
341 /* And zero out the PGD entry we we never release it twice. */
210 spgd->raw.val = 0; 342 spgd->raw.val = 0;
211 } 343 }
212} 344}
213 345
346/*H:440 (v) Flushing (thowing away) page tables,
347 *
348 * We saw flush_user_mappings() called when we re-used a top-level pgdir page.
349 * It simply releases every PTE page from 0 up to the kernel address. */
214static void flush_user_mappings(struct lguest *lg, int idx) 350static void flush_user_mappings(struct lguest *lg, int idx)
215{ 351{
216 unsigned int i; 352 unsigned int i;
353 /* Release every pgd entry up to the kernel's address. */
217 for (i = 0; i < vaddr_to_pgd_index(lg->page_offset); i++) 354 for (i = 0; i < vaddr_to_pgd_index(lg->page_offset); i++)
218 release_pgd(lg, lg->pgdirs[idx].pgdir + i); 355 release_pgd(lg, lg->pgdirs[idx].pgdir + i);
219} 356}
220 357
358/* The Guest also has a hypercall to do this manually: it's used when a large
359 * number of mappings have been changed. */
221void guest_pagetable_flush_user(struct lguest *lg) 360void guest_pagetable_flush_user(struct lguest *lg)
222{ 361{
362 /* Drop the userspace part of the current page table. */
223 flush_user_mappings(lg, lg->pgdidx); 363 flush_user_mappings(lg, lg->pgdidx);
224} 364}
365/*:*/
225 366
367/* We keep several page tables. This is a simple routine to find the page
368 * table (if any) corresponding to this top-level address the Guest has given
369 * us. */
226static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) 370static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable)
227{ 371{
228 unsigned int i; 372 unsigned int i;
@@ -232,21 +376,30 @@ static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable)
232 return i; 376 return i;
233} 377}
234 378
379/*H:435 And this is us, creating the new page directory. If we really do
380 * allocate a new one (and so the kernel parts are not there), we set
381 * blank_pgdir. */
235static unsigned int new_pgdir(struct lguest *lg, 382static unsigned int new_pgdir(struct lguest *lg,
236 unsigned long cr3, 383 unsigned long cr3,
237 int *blank_pgdir) 384 int *blank_pgdir)
238{ 385{
239 unsigned int next; 386 unsigned int next;
240 387
388 /* We pick one entry at random to throw out. Choosing the Least
389 * Recently Used might be better, but this is easy. */
241 next = random32() % ARRAY_SIZE(lg->pgdirs); 390 next = random32() % ARRAY_SIZE(lg->pgdirs);
391 /* If it's never been allocated at all before, try now. */
242 if (!lg->pgdirs[next].pgdir) { 392 if (!lg->pgdirs[next].pgdir) {
243 lg->pgdirs[next].pgdir = (spgd_t *)get_zeroed_page(GFP_KERNEL); 393 lg->pgdirs[next].pgdir = (spgd_t *)get_zeroed_page(GFP_KERNEL);
394 /* If the allocation fails, just keep using the one we have */
244 if (!lg->pgdirs[next].pgdir) 395 if (!lg->pgdirs[next].pgdir)
245 next = lg->pgdidx; 396 next = lg->pgdidx;
246 else 397 else
247 /* There are no mappings: you'll need to re-pin */ 398 /* This is a blank page, so there are no kernel
399 * mappings: caller must map the stack! */
248 *blank_pgdir = 1; 400 *blank_pgdir = 1;
249 } 401 }
402 /* Record which Guest toplevel this shadows. */
250 lg->pgdirs[next].cr3 = cr3; 403 lg->pgdirs[next].cr3 = cr3;
251 /* Release all the non-kernel mappings. */ 404 /* Release all the non-kernel mappings. */
252 flush_user_mappings(lg, next); 405 flush_user_mappings(lg, next);
@@ -254,82 +407,161 @@ static unsigned int new_pgdir(struct lguest *lg,
254 return next; 407 return next;
255} 408}
256 409
410/*H:430 (iv) Switching page tables
411 *
412 * This is what happens when the Guest changes page tables (ie. changes the
413 * top-level pgdir). This happens on almost every context switch. */
257void guest_new_pagetable(struct lguest *lg, unsigned long pgtable) 414void guest_new_pagetable(struct lguest *lg, unsigned long pgtable)
258{ 415{
259 int newpgdir, repin = 0; 416 int newpgdir, repin = 0;
260 417
418 /* Look to see if we have this one already. */
261 newpgdir = find_pgdir(lg, pgtable); 419 newpgdir = find_pgdir(lg, pgtable);
420 /* If not, we allocate or mug an existing one: if it's a fresh one,
421 * repin gets set to 1. */
262 if (newpgdir == ARRAY_SIZE(lg->pgdirs)) 422 if (newpgdir == ARRAY_SIZE(lg->pgdirs))
263 newpgdir = new_pgdir(lg, pgtable, &repin); 423 newpgdir = new_pgdir(lg, pgtable, &repin);
424 /* Change the current pgd index to the new one. */
264 lg->pgdidx = newpgdir; 425 lg->pgdidx = newpgdir;
426 /* If it was completely blank, we map in the Guest kernel stack */
265 if (repin) 427 if (repin)
266 pin_stack_pages(lg); 428 pin_stack_pages(lg);
267} 429}
268 430
431/*H:470 Finally, a routine which throws away everything: all PGD entries in all
432 * the shadow page tables. This is used when we destroy the Guest. */
269static void release_all_pagetables(struct lguest *lg) 433static void release_all_pagetables(struct lguest *lg)
270{ 434{
271 unsigned int i, j; 435 unsigned int i, j;
272 436
437 /* Every shadow pagetable this Guest has */
273 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) 438 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
274 if (lg->pgdirs[i].pgdir) 439 if (lg->pgdirs[i].pgdir)
440 /* Every PGD entry except the Switcher at the top */
275 for (j = 0; j < SWITCHER_PGD_INDEX; j++) 441 for (j = 0; j < SWITCHER_PGD_INDEX; j++)
276 release_pgd(lg, lg->pgdirs[i].pgdir + j); 442 release_pgd(lg, lg->pgdirs[i].pgdir + j);
277} 443}
278 444
445/* We also throw away everything when a Guest tells us it's changed a kernel
446 * mapping. Since kernel mappings are in every page table, it's easiest to
447 * throw them all away. This is amazingly slow, but thankfully rare. */
279void guest_pagetable_clear_all(struct lguest *lg) 448void guest_pagetable_clear_all(struct lguest *lg)
280{ 449{
281 release_all_pagetables(lg); 450 release_all_pagetables(lg);
451 /* We need the Guest kernel stack mapped again. */
282 pin_stack_pages(lg); 452 pin_stack_pages(lg);
283} 453}
284 454
455/*H:420 This is the routine which actually sets the page table entry for then
456 * "idx"'th shadow page table.
457 *
458 * Normally, we can just throw out the old entry and replace it with 0: if they
459 * use it demand_page() will put the new entry in. We need to do this anyway:
460 * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
461 * is read from, and _PAGE_DIRTY when it's written to.
462 *
463 * But Avi Kivity pointed out that most Operating Systems (Linux included) set
464 * these bits on PTEs immediately anyway. This is done to save the CPU from
465 * having to update them, but it helps us the same way: if they set
466 * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
467 * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
468 */
285static void do_set_pte(struct lguest *lg, int idx, 469static void do_set_pte(struct lguest *lg, int idx,
286 unsigned long vaddr, gpte_t gpte) 470 unsigned long vaddr, gpte_t gpte)
287{ 471{
472 /* Look up the matching shadow page directot entry. */
288 spgd_t *spgd = spgd_addr(lg, idx, vaddr); 473 spgd_t *spgd = spgd_addr(lg, idx, vaddr);
474
475 /* If the top level isn't present, there's no entry to update. */
289 if (spgd->flags & _PAGE_PRESENT) { 476 if (spgd->flags & _PAGE_PRESENT) {
477 /* Otherwise, we start by releasing the existing entry. */
290 spte_t *spte = spte_addr(lg, *spgd, vaddr); 478 spte_t *spte = spte_addr(lg, *spgd, vaddr);
291 release_pte(*spte); 479 release_pte(*spte);
480
481 /* If they're setting this entry as dirty or accessed, we might
482 * as well put that entry they've given us in now. This shaves
483 * 10% off a copy-on-write micro-benchmark. */
292 if (gpte.flags & (_PAGE_DIRTY | _PAGE_ACCESSED)) { 484 if (gpte.flags & (_PAGE_DIRTY | _PAGE_ACCESSED)) {
293 check_gpte(lg, gpte); 485 check_gpte(lg, gpte);
294 *spte = gpte_to_spte(lg, gpte, gpte.flags&_PAGE_DIRTY); 486 *spte = gpte_to_spte(lg, gpte, gpte.flags&_PAGE_DIRTY);
295 } else 487 } else
488 /* Otherwise we can demand_page() it in later. */
296 spte->raw.val = 0; 489 spte->raw.val = 0;
297 } 490 }
298} 491}
299 492
493/*H:410 Updating a PTE entry is a little trickier.
494 *
495 * We keep track of several different page tables (the Guest uses one for each
496 * process, so it makes sense to cache at least a few). Each of these have
497 * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
498 * all processes. So when the page table above that address changes, we update
499 * all the page tables, not just the current one. This is rare.
500 *
501 * The benefit is that when we have to track a new page table, we can copy keep
502 * all the kernel mappings. This speeds up context switch immensely. */
300void guest_set_pte(struct lguest *lg, 503void guest_set_pte(struct lguest *lg,
301 unsigned long cr3, unsigned long vaddr, gpte_t gpte) 504 unsigned long cr3, unsigned long vaddr, gpte_t gpte)
302{ 505{
303 /* Kernel mappings must be changed on all top levels. */ 506 /* Kernel mappings must be changed on all top levels. Slow, but
507 * doesn't happen often. */
304 if (vaddr >= lg->page_offset) { 508 if (vaddr >= lg->page_offset) {
305 unsigned int i; 509 unsigned int i;
306 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) 510 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
307 if (lg->pgdirs[i].pgdir) 511 if (lg->pgdirs[i].pgdir)
308 do_set_pte(lg, i, vaddr, gpte); 512 do_set_pte(lg, i, vaddr, gpte);
309 } else { 513 } else {
514 /* Is this page table one we have a shadow for? */
310 int pgdir = find_pgdir(lg, cr3); 515 int pgdir = find_pgdir(lg, cr3);
311 if (pgdir != ARRAY_SIZE(lg->pgdirs)) 516 if (pgdir != ARRAY_SIZE(lg->pgdirs))
517 /* If so, do the update. */
312 do_set_pte(lg, pgdir, vaddr, gpte); 518 do_set_pte(lg, pgdir, vaddr, gpte);
313 } 519 }
314} 520}
315 521
522/*H:400
523 * (iii) Setting up a page table entry when the Guest tells us it has changed.
524 *
525 * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
526 * with the other side of page tables while we're here: what happens when the
527 * Guest asks for a page table to be updated?
528 *
529 * We already saw that demand_page() will fill in the shadow page tables when
530 * needed, so we can simply remove shadow page table entries whenever the Guest
531 * tells us they've changed. When the Guest tries to use the new entry it will
532 * fault and demand_page() will fix it up.
533 *
534 * So with that in mind here's our code to to update a (top-level) PGD entry:
535 */
316void guest_set_pmd(struct lguest *lg, unsigned long cr3, u32 idx) 536void guest_set_pmd(struct lguest *lg, unsigned long cr3, u32 idx)
317{ 537{
318 int pgdir; 538 int pgdir;
319 539
540 /* The kernel seems to try to initialize this early on: we ignore its
541 * attempts to map over the Switcher. */
320 if (idx >= SWITCHER_PGD_INDEX) 542 if (idx >= SWITCHER_PGD_INDEX)
321 return; 543 return;
322 544
545 /* If they're talking about a page table we have a shadow for... */
323 pgdir = find_pgdir(lg, cr3); 546 pgdir = find_pgdir(lg, cr3);
324 if (pgdir < ARRAY_SIZE(lg->pgdirs)) 547 if (pgdir < ARRAY_SIZE(lg->pgdirs))
548 /* ... throw it away. */
325 release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx); 549 release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx);
326} 550}
327 551
552/*H:500 (vii) Setting up the page tables initially.
553 *
554 * When a Guest is first created, the Launcher tells us where the toplevel of
555 * its first page table is. We set some things up here: */
328int init_guest_pagetable(struct lguest *lg, unsigned long pgtable) 556int init_guest_pagetable(struct lguest *lg, unsigned long pgtable)
329{ 557{
330 /* We assume this in flush_user_mappings, so check now */ 558 /* In flush_user_mappings() we loop from 0 to
559 * "vaddr_to_pgd_index(lg->page_offset)". This assumes it won't hit
560 * the Switcher mappings, so check that now. */
331 if (vaddr_to_pgd_index(lg->page_offset) >= SWITCHER_PGD_INDEX) 561 if (vaddr_to_pgd_index(lg->page_offset) >= SWITCHER_PGD_INDEX)
332 return -EINVAL; 562 return -EINVAL;
563 /* We start on the first shadow page table, and give it a blank PGD
564 * page. */
333 lg->pgdidx = 0; 565 lg->pgdidx = 0;
334 lg->pgdirs[lg->pgdidx].cr3 = pgtable; 566 lg->pgdirs[lg->pgdidx].cr3 = pgtable;
335 lg->pgdirs[lg->pgdidx].pgdir = (spgd_t*)get_zeroed_page(GFP_KERNEL); 567 lg->pgdirs[lg->pgdidx].pgdir = (spgd_t*)get_zeroed_page(GFP_KERNEL);
@@ -338,33 +570,48 @@ int init_guest_pagetable(struct lguest *lg, unsigned long pgtable)
338 return 0; 570 return 0;
339} 571}
340 572
573/* When a Guest dies, our cleanup is fairly simple. */
341void free_guest_pagetable(struct lguest *lg) 574void free_guest_pagetable(struct lguest *lg)
342{ 575{
343 unsigned int i; 576 unsigned int i;
344 577
578 /* Throw away all page table pages. */
345 release_all_pagetables(lg); 579 release_all_pagetables(lg);
580 /* Now free the top levels: free_page() can handle 0 just fine. */
346 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) 581 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
347 free_page((long)lg->pgdirs[i].pgdir); 582 free_page((long)lg->pgdirs[i].pgdir);
348} 583}
349 584
350/* Caller must be preempt-safe */ 585/*H:480 (vi) Mapping the Switcher when the Guest is about to run.
586 *
587 * The Switcher and the two pages for this CPU need to be available to the
588 * Guest (and not the pages for other CPUs). We have the appropriate PTE pages
589 * for each CPU already set up, we just need to hook them in. */
351void map_switcher_in_guest(struct lguest *lg, struct lguest_pages *pages) 590void map_switcher_in_guest(struct lguest *lg, struct lguest_pages *pages)
352{ 591{
353 spte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages); 592 spte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages);
354 spgd_t switcher_pgd; 593 spgd_t switcher_pgd;
355 spte_t regs_pte; 594 spte_t regs_pte;
356 595
357 /* Since switcher less that 4MB, we simply mug top pte page. */ 596 /* Make the last PGD entry for this Guest point to the Switcher's PTE
597 * page for this CPU (with appropriate flags). */
358 switcher_pgd.pfn = __pa(switcher_pte_page) >> PAGE_SHIFT; 598 switcher_pgd.pfn = __pa(switcher_pte_page) >> PAGE_SHIFT;
359 switcher_pgd.flags = _PAGE_KERNEL; 599 switcher_pgd.flags = _PAGE_KERNEL;
360 lg->pgdirs[lg->pgdidx].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; 600 lg->pgdirs[lg->pgdidx].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd;
361 601
362 /* Map our regs page over stack page. */ 602 /* We also change the Switcher PTE page. When we're running the Guest,
603 * we want the Guest's "regs" page to appear where the first Switcher
604 * page for this CPU is. This is an optimization: when the Switcher
605 * saves the Guest registers, it saves them into the first page of this
606 * CPU's "struct lguest_pages": if we make sure the Guest's register
607 * page is already mapped there, we don't have to copy them out
608 * again. */
363 regs_pte.pfn = __pa(lg->regs_page) >> PAGE_SHIFT; 609 regs_pte.pfn = __pa(lg->regs_page) >> PAGE_SHIFT;
364 regs_pte.flags = _PAGE_KERNEL; 610 regs_pte.flags = _PAGE_KERNEL;
365 switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTES_PER_PAGE] 611 switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTES_PER_PAGE]
366 = regs_pte; 612 = regs_pte;
367} 613}
614/*:*/
368 615
369static void free_switcher_pte_pages(void) 616static void free_switcher_pte_pages(void)
370{ 617{
@@ -374,6 +621,10 @@ static void free_switcher_pte_pages(void)
374 free_page((long)switcher_pte_page(i)); 621 free_page((long)switcher_pte_page(i));
375} 622}
376 623
624/*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given
625 * the CPU number and the "struct page"s for the Switcher code itself.
626 *
627 * Currently the Switcher is less than a page long, so "pages" is always 1. */
377static __init void populate_switcher_pte_page(unsigned int cpu, 628static __init void populate_switcher_pte_page(unsigned int cpu,
378 struct page *switcher_page[], 629 struct page *switcher_page[],
379 unsigned int pages) 630 unsigned int pages)
@@ -381,21 +632,26 @@ static __init void populate_switcher_pte_page(unsigned int cpu,
381 unsigned int i; 632 unsigned int i;
382 spte_t *pte = switcher_pte_page(cpu); 633 spte_t *pte = switcher_pte_page(cpu);
383 634
635 /* The first entries are easy: they map the Switcher code. */
384 for (i = 0; i < pages; i++) { 636 for (i = 0; i < pages; i++) {
385 pte[i].pfn = page_to_pfn(switcher_page[i]); 637 pte[i].pfn = page_to_pfn(switcher_page[i]);
386 pte[i].flags = _PAGE_PRESENT|_PAGE_ACCESSED; 638 pte[i].flags = _PAGE_PRESENT|_PAGE_ACCESSED;
387 } 639 }
388 640
389 /* We only map this CPU's pages, so guest can't see others. */ 641 /* The only other thing we map is this CPU's pair of pages. */
390 i = pages + cpu*2; 642 i = pages + cpu*2;
391 643
392 /* First page (regs) is rw, second (state) is ro. */ 644 /* First page (Guest registers) is writable from the Guest */
393 pte[i].pfn = page_to_pfn(switcher_page[i]); 645 pte[i].pfn = page_to_pfn(switcher_page[i]);
394 pte[i].flags = _PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW; 646 pte[i].flags = _PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW;
647 /* The second page contains the "struct lguest_ro_state", and is
648 * read-only. */
395 pte[i+1].pfn = page_to_pfn(switcher_page[i+1]); 649 pte[i+1].pfn = page_to_pfn(switcher_page[i+1]);
396 pte[i+1].flags = _PAGE_PRESENT|_PAGE_ACCESSED; 650 pte[i+1].flags = _PAGE_PRESENT|_PAGE_ACCESSED;
397} 651}
398 652
653/*H:510 At boot or module load time, init_pagetables() allocates and populates
654 * the Switcher PTE page for each CPU. */
399__init int init_pagetables(struct page **switcher_page, unsigned int pages) 655__init int init_pagetables(struct page **switcher_page, unsigned int pages)
400{ 656{
401 unsigned int i; 657 unsigned int i;
@@ -410,7 +666,9 @@ __init int init_pagetables(struct page **switcher_page, unsigned int pages)
410 } 666 }
411 return 0; 667 return 0;
412} 668}
669/*:*/
413 670
671/* Cleaning up simply involves freeing the PTE page for each CPU. */
414void free_pagetables(void) 672void free_pagetables(void)
415{ 673{
416 free_switcher_pte_pages(); 674 free_switcher_pte_pages();