blob: a8d0aee3bc0e3dfa7943cd412137a0d44e2d737d [file] [log] [blame]
Rusty Russell2e04ef72009-07-30 16:03:45 -06001/*P:700
2 * The pagetable code, on the other hand, still shows the scars of
Rusty Russellf938d2c2007-07-26 10:41:02 -07003 * 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
Rusty Russella6bd8e12008-03-28 11:05:53 -05006 * it nor use it: we verify and convert it here then point the CPU to the
Rusty Russell2e04ef72009-07-30 16:03:45 -06007 * converted Guest pages when running the Guest.
8:*/
Rusty Russellf938d2c2007-07-26 10:41:02 -07009
10/* Copyright (C) Rusty Russell IBM Corporation 2006.
Rusty Russelld7e28ff2007-07-19 01:49:23 -070011 * GPL v2 and any later version */
12#include <linux/mm.h>
13#include <linux/types.h>
14#include <linux/spinlock.h>
15#include <linux/random.h>
16#include <linux/percpu.h>
17#include <asm/tlbflush.h>
Rusty Russell47436aa2007-10-22 11:03:36 +100018#include <asm/uaccess.h>
Matias Zabaljauregui58a24562008-09-29 01:40:07 -030019#include <asm/bootparam.h>
Rusty Russelld7e28ff2007-07-19 01:49:23 -070020#include "lg.h"
21
Rusty Russell2e04ef72009-07-30 16:03:45 -060022/*M:008
23 * We hold reference to pages, which prevents them from being swapped.
Rusty Russellf56a3842007-07-26 10:41:05 -070024 * 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
Rusty Russell2e04ef72009-07-30 16:03:45 -060026 * could probably consider launching Guests as non-root.
27:*/
Rusty Russellf56a3842007-07-26 10:41:05 -070028
Rusty Russellbff672e2007-07-26 10:41:04 -070029/*H:300
30 * The Page Table Code
31 *
Rusty Russella91d74a2009-07-30 16:03:45 -060032 * 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!).
Rusty Russellbff672e2007-07-26 10:41:04 -070036 *
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:
Rusty Russelle1e72962007-10-25 15:02:50 +100045 * (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,
Rusty Russellbff672e2007-07-26 10:41:04 -070048 * (iv) Switching page tables,
Rusty Russelle1e72962007-10-25 15:02:50 +100049 * (v) Flushing (throwing away) page tables,
Rusty Russellbff672e2007-07-26 10:41:04 -070050 * (vi) Mapping the Switcher when the Guest is about to run,
51 * (vii) Setting up the page tables initially.
Rusty Russell2e04ef72009-07-30 16:03:45 -060052:*/
Rusty Russellbff672e2007-07-26 10:41:04 -070053
Rusty Russell2e04ef72009-07-30 16:03:45 -060054/*
Rusty Russella91d74a2009-07-30 16:03:45 -060055 * The Switcher uses the complete top PTE page. That's 1024 PTE entries (4MB)
56 * or 512 PTE entries with PAE (2MB).
Rusty Russell2e04ef72009-07-30 16:03:45 -060057 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +100058#define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1)
Rusty Russelld7e28ff2007-07-19 01:49:23 -070059
Rusty Russell2e04ef72009-07-30 16:03:45 -060060/*
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 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -060064#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
Rusty Russell2e04ef72009-07-30 16:03:45 -060073/*
74 * We actually need a separate PTE page for each CPU. Remember that after the
Rusty Russellbff672e2007-07-26 10:41:04 -070075 * Switcher code itself comes two pages for each CPU, and we don't want this
Rusty Russell2e04ef72009-07-30 16:03:45 -060076 * CPU's guest to see the pages of any other CPU.
77 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +100078static DEFINE_PER_CPU(pte_t *, switcher_pte_pages);
Rusty Russelld7e28ff2007-07-19 01:49:23 -070079#define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
80
Rusty Russell2e04ef72009-07-30 16:03:45 -060081/*H:320
82 * The page table code is curly enough to need helper functions to keep it
Rusty Russella91d74a2009-07-30 16:03:45 -060083 * 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.
Rusty Russellbff672e2007-07-26 10:41:04 -070085 *
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +100086 * There are two functions which return pointers to the shadow (aka "real")
Rusty Russellbff672e2007-07-26 10:41:04 -070087 * page tables.
88 *
89 * spgd_addr() takes the virtual address and returns a pointer to the top-level
Rusty Russelle1e72962007-10-25 15:02:50 +100090 * 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
Rusty Russell2e04ef72009-07-30 16:03:45 -060092 * usually the current one).
93 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -020094static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr)
Rusty Russelld7e28ff2007-07-19 01:49:23 -070095{
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +100096 unsigned int index = pgd_index(vaddr);
Rusty Russelld7e28ff2007-07-19 01:49:23 -070097
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -060098#ifndef CONFIG_X86_PAE
Rusty Russellbff672e2007-07-26 10:41:04 -070099 /* We kill any Guest trying to touch the Switcher addresses. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700100 if (index >= SWITCHER_PGD_INDEX) {
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200101 kill_guest(cpu, "attempt to access switcher pages");
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700102 index = 0;
103 }
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600104#endif
Rusty Russellbff672e2007-07-26 10:41:04 -0700105 /* Return a pointer index'th pgd entry for the i'th page table. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200106 return &cpu->lg->pgdirs[i].pgdir[index];
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700107}
108
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600109#ifdef CONFIG_X86_PAE
Rusty Russell2e04ef72009-07-30 16:03:45 -0600110/*
111 * This routine then takes the PGD entry given above, which contains the
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600112 * address of the PMD page. It then returns a pointer to the PMD entry for the
Rusty Russell2e04ef72009-07-30 16:03:45 -0600113 * given address.
114 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600115static pmd_t *spmd_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr)
116{
117 unsigned int index = pmd_index(vaddr);
118 pmd_t *page;
119
120 /* We kill any Guest trying to touch the Switcher addresses. */
121 if (pgd_index(vaddr) == SWITCHER_PGD_INDEX &&
122 index >= SWITCHER_PMD_INDEX) {
123 kill_guest(cpu, "attempt to access switcher pages");
124 index = 0;
125 }
126
127 /* You should never call this if the PGD entry wasn't valid */
128 BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT));
129 page = __va(pgd_pfn(spgd) << PAGE_SHIFT);
130
131 return &page[index];
132}
133#endif
134
Rusty Russell2e04ef72009-07-30 16:03:45 -0600135/*
136 * This routine then takes the page directory entry returned above, which
Rusty Russelle1e72962007-10-25 15:02:50 +1000137 * contains the address of the page table entry (PTE) page. It then returns a
Rusty Russell2e04ef72009-07-30 16:03:45 -0600138 * pointer to the PTE entry for the given address.
139 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600140static pte_t *spte_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700141{
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600142#ifdef CONFIG_X86_PAE
143 pmd_t *pmd = spmd_addr(cpu, spgd, vaddr);
144 pte_t *page = __va(pmd_pfn(*pmd) << PAGE_SHIFT);
145
146 /* You should never call this if the PMD entry wasn't valid */
147 BUG_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT));
148#else
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000149 pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT);
Rusty Russellbff672e2007-07-26 10:41:04 -0700150 /* You should never call this if the PGD entry wasn't valid */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000151 BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT));
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600152#endif
153
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600154 return &page[pte_index(vaddr)];
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700155}
156
Rusty Russell2e04ef72009-07-30 16:03:45 -0600157/*
Rusty Russella91d74a2009-07-30 16:03:45 -0600158 * These functions are just like the above two, except they access the Guest
Rusty Russell2e04ef72009-07-30 16:03:45 -0600159 * page tables. Hence they return a Guest address.
160 */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200161static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700162{
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000163 unsigned int index = vaddr >> (PGDIR_SHIFT);
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200164 return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700165}
166
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600167#ifdef CONFIG_X86_PAE
Rusty Russella91d74a2009-07-30 16:03:45 -0600168/* Follow the PGD to the PMD. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600169static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700170{
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000171 unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT;
172 BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT));
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600173 return gpage + pmd_index(vaddr) * sizeof(pmd_t);
174}
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600175
Rusty Russella91d74a2009-07-30 16:03:45 -0600176/* Follow the PMD to the PTE. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600177static unsigned long gpte_addr(struct lg_cpu *cpu,
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600178 pmd_t gpmd, unsigned long vaddr)
179{
180 unsigned long gpage = pmd_pfn(gpmd) << PAGE_SHIFT;
181
182 BUG_ON(!(pmd_flags(gpmd) & _PAGE_PRESENT));
183 return gpage + pte_index(vaddr) * sizeof(pte_t);
184}
185#else
Rusty Russella91d74a2009-07-30 16:03:45 -0600186/* Follow the PGD to the PTE (no mid-level for !PAE). */
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600187static unsigned long gpte_addr(struct lg_cpu *cpu,
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600188 pgd_t gpgd, unsigned long vaddr)
189{
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600190 unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT;
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600191
192 BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT));
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600193 return gpage + pte_index(vaddr) * sizeof(pte_t);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700194}
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600195#endif
Rusty Russella6bd8e12008-03-28 11:05:53 -0500196/*:*/
197
Rusty Russell2e04ef72009-07-30 16:03:45 -0600198/*M:014
199 * get_pfn is slow: we could probably try to grab batches of pages here as
200 * an optimization (ie. pre-faulting).
201:*/
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700202
Rusty Russell2e04ef72009-07-30 16:03:45 -0600203/*H:350
204 * This routine takes a page number given by the Guest and converts it to
Rusty Russellbff672e2007-07-26 10:41:04 -0700205 * an actual, physical page number. It can fail for several reasons: the
206 * virtual address might not be mapped by the Launcher, the write flag is set
207 * and the page is read-only, or the write flag was set and the page was
208 * shared so had to be copied, but we ran out of memory.
209 *
Rusty Russella6bd8e12008-03-28 11:05:53 -0500210 * This holds a reference to the page, so release_pte() is careful to put that
Rusty Russell2e04ef72009-07-30 16:03:45 -0600211 * back.
212 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700213static unsigned long get_pfn(unsigned long virtpfn, int write)
214{
215 struct page *page;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700216
Rusty Russell71a3f4e2008-08-12 17:52:53 -0500217 /* gup me one page at this address please! */
218 if (get_user_pages_fast(virtpfn << PAGE_SHIFT, 1, write, &page) == 1)
219 return page_to_pfn(page);
220
221 /* This value indicates failure. */
222 return -1UL;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700223}
224
Rusty Russell2e04ef72009-07-30 16:03:45 -0600225/*H:340
226 * Converting a Guest page table entry to a shadow (ie. real) page table
Rusty Russellbff672e2007-07-26 10:41:04 -0700227 * entry can be a little tricky. The flags are (almost) the same, but the
228 * Guest PTE contains a virtual page number: the CPU needs the real page
Rusty Russell2e04ef72009-07-30 16:03:45 -0600229 * number.
230 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200231static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700232{
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000233 unsigned long pfn, base, flags;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700234
Rusty Russell2e04ef72009-07-30 16:03:45 -0600235 /*
236 * The Guest sets the global flag, because it thinks that it is using
Rusty Russellbff672e2007-07-26 10:41:04 -0700237 * PGE. We only told it to use PGE so it would tell us whether it was
238 * flushing a kernel mapping or a userspace mapping. We don't actually
Rusty Russell2e04ef72009-07-30 16:03:45 -0600239 * use the global bit, so throw it away.
240 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000241 flags = (pte_flags(gpte) & ~_PAGE_GLOBAL);
Rusty Russellbff672e2007-07-26 10:41:04 -0700242
Rusty Russell3c6b5bf2007-10-22 11:03:26 +1000243 /* The Guest's pages are offset inside the Launcher. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200244 base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE;
Rusty Russell3c6b5bf2007-10-22 11:03:26 +1000245
Rusty Russell2e04ef72009-07-30 16:03:45 -0600246 /*
247 * We need a temporary "unsigned long" variable to hold the answer from
Rusty Russellbff672e2007-07-26 10:41:04 -0700248 * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
249 * fit in spte.pfn. get_pfn() finds the real physical number of the
Rusty Russell2e04ef72009-07-30 16:03:45 -0600250 * page, given the virtual number.
251 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000252 pfn = get_pfn(base + pte_pfn(gpte), write);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700253 if (pfn == -1UL) {
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200254 kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte));
Rusty Russell2e04ef72009-07-30 16:03:45 -0600255 /*
256 * When we destroy the Guest, we'll go through the shadow page
Rusty Russellbff672e2007-07-26 10:41:04 -0700257 * tables and release_pte() them. Make sure we don't think
Rusty Russell2e04ef72009-07-30 16:03:45 -0600258 * this one is valid!
259 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000260 flags = 0;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700261 }
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000262 /* Now we assemble our shadow PTE from the page number and flags. */
263 return pfn_pte(pfn, __pgprot(flags));
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700264}
265
Rusty Russellbff672e2007-07-26 10:41:04 -0700266/*H:460 And to complete the chain, release_pte() looks like this: */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000267static void release_pte(pte_t pte)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700268{
Rusty Russell2e04ef72009-07-30 16:03:45 -0600269 /*
270 * Remember that get_user_pages_fast() took a reference to the page, in
271 * get_pfn()? We have to put it back now.
272 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000273 if (pte_flags(pte) & _PAGE_PRESENT)
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600274 put_page(pte_page(pte));
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700275}
Rusty Russellbff672e2007-07-26 10:41:04 -0700276/*:*/
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700277
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200278static void check_gpte(struct lg_cpu *cpu, pte_t gpte)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700279{
Ahmed S. Darwish31f4b462008-02-09 23:24:09 +0100280 if ((pte_flags(gpte) & _PAGE_PSE) ||
281 pte_pfn(gpte) >= cpu->lg->pfn_limit)
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200282 kill_guest(cpu, "bad page table entry");
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700283}
284
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200285static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700286{
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600287 if ((pgd_flags(gpgd) & ~CHECK_GPGD_MASK) ||
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200288 (pgd_pfn(gpgd) >= cpu->lg->pfn_limit))
289 kill_guest(cpu, "bad page directory entry");
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700290}
291
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600292#ifdef CONFIG_X86_PAE
293static void check_gpmd(struct lg_cpu *cpu, pmd_t gpmd)
294{
295 if ((pmd_flags(gpmd) & ~_PAGE_TABLE) ||
296 (pmd_pfn(gpmd) >= cpu->lg->pfn_limit))
297 kill_guest(cpu, "bad page middle directory entry");
298}
299#endif
300
Rusty Russellbff672e2007-07-26 10:41:04 -0700301/*H:330
Rusty Russelle1e72962007-10-25 15:02:50 +1000302 * (i) Looking up a page table entry when the Guest faults.
Rusty Russellbff672e2007-07-26 10:41:04 -0700303 *
304 * We saw this call in run_guest(): when we see a page fault in the Guest, we
305 * come here. That's because we only set up the shadow page tables lazily as
306 * they're needed, so we get page faults all the time and quietly fix them up
307 * and return to the Guest without it knowing.
308 *
309 * If we fixed up the fault (ie. we mapped the address), this routine returns
Rusty Russell2e04ef72009-07-30 16:03:45 -0600310 * true. Otherwise, it was a real fault and we need to tell the Guest.
311 */
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300312bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700313{
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000314 pgd_t gpgd;
315 pgd_t *spgd;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700316 unsigned long gpte_ptr;
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000317 pte_t gpte;
318 pte_t *spte;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700319
Rusty Russella91d74a2009-07-30 16:03:45 -0600320 /* Mid level for PAE. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600321#ifdef CONFIG_X86_PAE
322 pmd_t *spmd;
323 pmd_t gpmd;
324#endif
325
Rusty Russellbff672e2007-07-26 10:41:04 -0700326 /* First step: get the top-level Guest page table entry. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200327 gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
Rusty Russellbff672e2007-07-26 10:41:04 -0700328 /* Toplevel not present? We can't map it in. */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000329 if (!(pgd_flags(gpgd) & _PAGE_PRESENT))
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300330 return false;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700331
Rusty Russellbff672e2007-07-26 10:41:04 -0700332 /* Now look at the matching shadow entry. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200333 spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr);
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000334 if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) {
Rusty Russellbff672e2007-07-26 10:41:04 -0700335 /* No shadow entry: allocate a new shadow PTE page. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700336 unsigned long ptepage = get_zeroed_page(GFP_KERNEL);
Rusty Russell2e04ef72009-07-30 16:03:45 -0600337 /*
338 * This is not really the Guest's fault, but killing it is
339 * simple for this corner case.
340 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700341 if (!ptepage) {
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200342 kill_guest(cpu, "out of memory allocating pte page");
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300343 return false;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700344 }
Rusty Russellbff672e2007-07-26 10:41:04 -0700345 /* We check that the Guest pgd is OK. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200346 check_gpgd(cpu, gpgd);
Rusty Russell2e04ef72009-07-30 16:03:45 -0600347 /*
348 * And we copy the flags to the shadow PGD entry. The page
349 * number in the shadow PGD is the page we just allocated.
350 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600351 set_pgd(spgd, __pgd(__pa(ptepage) | pgd_flags(gpgd)));
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700352 }
353
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600354#ifdef CONFIG_X86_PAE
355 gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t);
Rusty Russell2e04ef72009-07-30 16:03:45 -0600356 /* Middle level not present? We can't map it in. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600357 if (!(pmd_flags(gpmd) & _PAGE_PRESENT))
358 return false;
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
Rusty Russell2e04ef72009-07-30 16:03:45 -0600367 /*
368 * This is not really the Guest's fault, but killing it is
369 * simple for this corner case.
370 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600371 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
Rusty Russell2e04ef72009-07-30 16:03:45 -0600379 /*
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 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600383 native_set_pmd(spmd, __pmd(__pa(ptepage) | pmd_flags(gpmd)));
384 }
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600385
Rusty Russell2e04ef72009-07-30 16:03:45 -0600386 /*
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 */
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600390 gpte_ptr = gpte_addr(cpu, gpmd, vaddr);
391#else
Rusty Russell2e04ef72009-07-30 16:03:45 -0600392 /*
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 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600396 gpte_ptr = gpte_addr(cpu, gpgd, vaddr);
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600397#endif
Rusty Russella91d74a2009-07-30 16:03:45 -0600398
399 /* Read the actual PTE value. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200400 gpte = lgread(cpu, gpte_ptr, pte_t);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700401
Rusty Russellbff672e2007-07-26 10:41:04 -0700402 /* If this page isn't in the Guest page tables, we can't page it in. */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000403 if (!(pte_flags(gpte) & _PAGE_PRESENT))
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300404 return false;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700405
Rusty Russell2e04ef72009-07-30 16:03:45 -0600406 /*
407 * Check they're not trying to write to a page the Guest wants
408 * read-only (bit 2 of errcode == write).
409 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000410 if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW))
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300411 return false;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700412
Rusty Russelle1e72962007-10-25 15:02:50 +1000413 /* User access to a kernel-only page? (bit 3 == user access) */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000414 if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER))
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300415 return false;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700416
Rusty Russell2e04ef72009-07-30 16:03:45 -0600417 /*
418 * Check that the Guest PTE flags are OK, and the page number is below
419 * the pfn_limit (ie. not mapping the Launcher binary).
420 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200421 check_gpte(cpu, gpte);
Rusty Russelle1e72962007-10-25 15:02:50 +1000422
Rusty Russellbff672e2007-07-26 10:41:04 -0700423 /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000424 gpte = pte_mkyoung(gpte);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700425 if (errcode & 2)
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000426 gpte = pte_mkdirty(gpte);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700427
Rusty Russellbff672e2007-07-26 10:41:04 -0700428 /* Get the pointer to the shadow PTE entry we're going to set. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600429 spte = spte_addr(cpu, *spgd, vaddr);
Rusty Russell2e04ef72009-07-30 16:03:45 -0600430
431 /*
432 * If there was a valid shadow PTE entry here before, we release it.
433 * This can happen with a write to a previously read-only entry.
434 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700435 release_pte(*spte);
436
Rusty Russell2e04ef72009-07-30 16:03:45 -0600437 /*
438 * If this is a write, we insist that the Guest page is writable (the
439 * final arg to gpte_to_spte()).
440 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000441 if (pte_dirty(gpte))
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200442 *spte = gpte_to_spte(cpu, gpte, 1);
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000443 else
Rusty Russell2e04ef72009-07-30 16:03:45 -0600444 /*
445 * If this is a read, don't set the "writable" bit in the page
Rusty Russellbff672e2007-07-26 10:41:04 -0700446 * table entry, even if the Guest says it's writable. That way
Rusty Russelle1e72962007-10-25 15:02:50 +1000447 * we will come back here when a write does actually occur, so
Rusty Russell2e04ef72009-07-30 16:03:45 -0600448 * we can update the Guest's _PAGE_DIRTY flag.
449 */
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600450 native_set_pte(spte, gpte_to_spte(cpu, pte_wrprotect(gpte), 0));
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700451
Rusty Russell2e04ef72009-07-30 16:03:45 -0600452 /*
453 * Finally, we write the Guest PTE entry back: we've set the
454 * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags.
455 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200456 lgwrite(cpu, gpte_ptr, pte_t, gpte);
Rusty Russellbff672e2007-07-26 10:41:04 -0700457
Rusty Russell2e04ef72009-07-30 16:03:45 -0600458 /*
459 * The fault is fixed, the page table is populated, the mapping
Rusty Russelle1e72962007-10-25 15:02:50 +1000460 * manipulated, the result returned and the code complete. A small
461 * delay and a trace of alliteration are the only indications the Guest
Rusty Russell2e04ef72009-07-30 16:03:45 -0600462 * has that a page fault occurred at all.
463 */
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300464 return true;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700465}
466
Rusty Russelle1e72962007-10-25 15:02:50 +1000467/*H:360
468 * (ii) Making sure the Guest stack is mapped.
Rusty Russellbff672e2007-07-26 10:41:04 -0700469 *
Rusty Russelle1e72962007-10-25 15:02:50 +1000470 * Remember that direct traps into the Guest need a mapped Guest kernel stack.
471 * pin_stack_pages() calls us here: we could simply call demand_page(), but as
472 * we've seen that logic is quite long, and usually the stack pages are already
473 * mapped, so it's overkill.
Rusty Russellbff672e2007-07-26 10:41:04 -0700474 *
475 * This is a quick version which answers the question: is this virtual address
Rusty Russell2e04ef72009-07-30 16:03:45 -0600476 * mapped by the shadow page tables, and is it writable?
477 */
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300478static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700479{
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000480 pgd_t *spgd;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700481 unsigned long flags;
482
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600483#ifdef CONFIG_X86_PAE
484 pmd_t *spmd;
485#endif
Rusty Russelle1e72962007-10-25 15:02:50 +1000486 /* Look at the current top level entry: is it present? */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200487 spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr);
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000488 if (!(pgd_flags(*spgd) & _PAGE_PRESENT))
Matias Zabaljaureguidf1693a2009-03-18 13:38:35 -0300489 return false;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700490
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600491#ifdef CONFIG_X86_PAE
492 spmd = spmd_addr(cpu, *spgd, vaddr);
493 if (!(pmd_flags(*spmd) & _PAGE_PRESENT))
494 return false;
495#endif
496
Rusty Russell2e04ef72009-07-30 16:03:45 -0600497 /*
498 * Check the flags on the pte entry itself: it must be present and
499 * writable.
500 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600501 flags = pte_flags(*(spte_addr(cpu, *spgd, vaddr)));
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000502
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700503 return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW);
504}
505
Rusty Russell2e04ef72009-07-30 16:03:45 -0600506/*
507 * So, when pin_stack_pages() asks us to pin a page, we check if it's already
Rusty Russellbff672e2007-07-26 10:41:04 -0700508 * in the page tables, and if not, we call demand_page() with error code 2
Rusty Russell2e04ef72009-07-30 16:03:45 -0600509 * (meaning "write").
510 */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200511void pin_page(struct lg_cpu *cpu, unsigned long vaddr)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700512{
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200513 if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2))
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200514 kill_guest(cpu, "bad stack page %#lx", vaddr);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700515}
Rusty Russella91d74a2009-07-30 16:03:45 -0600516/*:*/
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700517
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600518#ifdef CONFIG_X86_PAE
519static void release_pmd(pmd_t *spmd)
520{
521 /* If the entry's not present, there's nothing to release. */
522 if (pmd_flags(*spmd) & _PAGE_PRESENT) {
523 unsigned int i;
524 pte_t *ptepage = __va(pmd_pfn(*spmd) << PAGE_SHIFT);
525 /* For each entry in the page, we might need to release it. */
526 for (i = 0; i < PTRS_PER_PTE; i++)
527 release_pte(ptepage[i]);
528 /* Now we can free the page of PTEs */
529 free_page((long)ptepage);
530 /* And zero out the PMD entry so we never release it twice. */
531 native_set_pmd(spmd, __pmd(0));
532 }
533}
534
535static void release_pgd(pgd_t *spgd)
536{
537 /* If the entry's not present, there's nothing to release. */
538 if (pgd_flags(*spgd) & _PAGE_PRESENT) {
539 unsigned int i;
540 pmd_t *pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
541
542 for (i = 0; i < PTRS_PER_PMD; i++)
543 release_pmd(&pmdpage[i]);
544
545 /* Now we can free the page of PMDs */
546 free_page((long)pmdpage);
547 /* And zero out the PGD entry so we never release it twice. */
548 set_pgd(spgd, __pgd(0));
549 }
550}
551
552#else /* !CONFIG_X86_PAE */
Rusty Russella91d74a2009-07-30 16:03:45 -0600553/*H:450
554 * If we chase down the release_pgd() code, the non-PAE version looks like
555 * this. The PAE version is almost identical, but instead of calling
556 * release_pte it calls release_pmd(), which looks much like this.
557 */
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600558static void release_pgd(pgd_t *spgd)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700559{
Rusty Russellbff672e2007-07-26 10:41:04 -0700560 /* If the entry's not present, there's nothing to release. */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000561 if (pgd_flags(*spgd) & _PAGE_PRESENT) {
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700562 unsigned int i;
Rusty Russell2e04ef72009-07-30 16:03:45 -0600563 /*
564 * Converting the pfn to find the actual PTE page is easy: turn
Rusty Russellbff672e2007-07-26 10:41:04 -0700565 * the page number into a physical address, then convert to a
Rusty Russell2e04ef72009-07-30 16:03:45 -0600566 * virtual address (easy for kernel pages like this one).
567 */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000568 pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
Rusty Russellbff672e2007-07-26 10:41:04 -0700569 /* For each entry in the page, we might need to release it. */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000570 for (i = 0; i < PTRS_PER_PTE; i++)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700571 release_pte(ptepage[i]);
Rusty Russellbff672e2007-07-26 10:41:04 -0700572 /* Now we can free the page of PTEs */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700573 free_page((long)ptepage);
Rusty Russelle1e72962007-10-25 15:02:50 +1000574 /* And zero out the PGD entry so we never release it twice. */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000575 *spgd = __pgd(0);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700576 }
577}
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600578#endif
Rusty Russell2e04ef72009-07-30 16:03:45 -0600579
580/*H:445
581 * We saw flush_user_mappings() twice: once from the flush_user_mappings()
Rusty Russelle1e72962007-10-25 15:02:50 +1000582 * hypercall and once in new_pgdir() when we re-used a top-level pgdir page.
Rusty Russell2e04ef72009-07-30 16:03:45 -0600583 * It simply releases every PTE page from 0 up to the Guest's kernel address.
584 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700585static void flush_user_mappings(struct lguest *lg, int idx)
586{
587 unsigned int i;
Rusty Russellbff672e2007-07-26 10:41:04 -0700588 /* Release every pgd entry up to the kernel's address. */
Rusty Russell47436aa2007-10-22 11:03:36 +1000589 for (i = 0; i < pgd_index(lg->kernel_address); i++)
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600590 release_pgd(lg->pgdirs[idx].pgdir + i);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700591}
592
Rusty Russell2e04ef72009-07-30 16:03:45 -0600593/*H:440
594 * (v) Flushing (throwing away) page tables,
Rusty Russelle1e72962007-10-25 15:02:50 +1000595 *
596 * The Guest has a hypercall to throw away the page tables: it's used when a
Rusty Russell2e04ef72009-07-30 16:03:45 -0600597 * large number of mappings have been changed.
598 */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200599void guest_pagetable_flush_user(struct lg_cpu *cpu)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700600{
Rusty Russellbff672e2007-07-26 10:41:04 -0700601 /* Drop the userspace part of the current page table. */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200602 flush_user_mappings(cpu->lg, cpu->cpu_pgd);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700603}
Rusty Russellbff672e2007-07-26 10:41:04 -0700604/*:*/
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700605
Rusty Russell47436aa2007-10-22 11:03:36 +1000606/* We walk down the guest page tables to get a guest-physical address */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200607unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr)
Rusty Russell47436aa2007-10-22 11:03:36 +1000608{
609 pgd_t gpgd;
610 pte_t gpte;
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600611#ifdef CONFIG_X86_PAE
612 pmd_t gpmd;
613#endif
Rusty Russell47436aa2007-10-22 11:03:36 +1000614 /* First step: get the top-level Guest page table entry. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200615 gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
Rusty Russell47436aa2007-10-22 11:03:36 +1000616 /* Toplevel not present? We can't map it in. */
Rusty Russell6afbdd02009-03-30 21:55:23 -0600617 if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) {
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200618 kill_guest(cpu, "Bad address %#lx", vaddr);
Rusty Russell6afbdd02009-03-30 21:55:23 -0600619 return -1UL;
620 }
Rusty Russell47436aa2007-10-22 11:03:36 +1000621
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600622#ifdef CONFIG_X86_PAE
623 gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t);
624 if (!(pmd_flags(gpmd) & _PAGE_PRESENT))
625 kill_guest(cpu, "Bad address %#lx", vaddr);
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600626 gpte = lgread(cpu, gpte_addr(cpu, gpmd, vaddr), pte_t);
627#else
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600628 gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t);
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600629#endif
Rusty Russell47436aa2007-10-22 11:03:36 +1000630 if (!(pte_flags(gpte) & _PAGE_PRESENT))
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200631 kill_guest(cpu, "Bad address %#lx", vaddr);
Rusty Russell47436aa2007-10-22 11:03:36 +1000632
633 return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK);
634}
635
Rusty Russell2e04ef72009-07-30 16:03:45 -0600636/*
637 * We keep several page tables. This is a simple routine to find the page
Rusty Russellbff672e2007-07-26 10:41:04 -0700638 * table (if any) corresponding to this top-level address the Guest has given
Rusty Russell2e04ef72009-07-30 16:03:45 -0600639 * us.
640 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700641static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable)
642{
643 unsigned int i;
644 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
Rusty Russell4357bd92008-03-11 09:35:57 -0500645 if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700646 break;
647 return i;
648}
649
Rusty Russell2e04ef72009-07-30 16:03:45 -0600650/*H:435
651 * And this is us, creating the new page directory. If we really do
Rusty Russellbff672e2007-07-26 10:41:04 -0700652 * allocate a new one (and so the kernel parts are not there), we set
Rusty Russell2e04ef72009-07-30 16:03:45 -0600653 * blank_pgdir.
654 */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200655static unsigned int new_pgdir(struct lg_cpu *cpu,
Rusty Russellee3db0f2007-10-22 11:03:34 +1000656 unsigned long gpgdir,
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700657 int *blank_pgdir)
658{
659 unsigned int next;
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600660#ifdef CONFIG_X86_PAE
661 pmd_t *pmd_table;
662#endif
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700663
Rusty Russell2e04ef72009-07-30 16:03:45 -0600664 /*
665 * We pick one entry at random to throw out. Choosing the Least
666 * Recently Used might be better, but this is easy.
667 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200668 next = random32() % ARRAY_SIZE(cpu->lg->pgdirs);
Rusty Russellbff672e2007-07-26 10:41:04 -0700669 /* If it's never been allocated at all before, try now. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200670 if (!cpu->lg->pgdirs[next].pgdir) {
671 cpu->lg->pgdirs[next].pgdir =
672 (pgd_t *)get_zeroed_page(GFP_KERNEL);
Rusty Russellbff672e2007-07-26 10:41:04 -0700673 /* If the allocation fails, just keep using the one we have */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200674 if (!cpu->lg->pgdirs[next].pgdir)
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200675 next = cpu->cpu_pgd;
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600676 else {
677#ifdef CONFIG_X86_PAE
Rusty Russell2e04ef72009-07-30 16:03:45 -0600678 /*
679 * In PAE mode, allocate a pmd page and populate the
680 * last pgd entry.
681 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600682 pmd_table = (pmd_t *)get_zeroed_page(GFP_KERNEL);
683 if (!pmd_table) {
684 free_page((long)cpu->lg->pgdirs[next].pgdir);
685 set_pgd(cpu->lg->pgdirs[next].pgdir, __pgd(0));
686 next = cpu->cpu_pgd;
687 } else {
688 set_pgd(cpu->lg->pgdirs[next].pgdir +
689 SWITCHER_PGD_INDEX,
690 __pgd(__pa(pmd_table) | _PAGE_PRESENT));
Rusty Russell2e04ef72009-07-30 16:03:45 -0600691 /*
692 * This is a blank page, so there are no kernel
693 * mappings: caller must map the stack!
694 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600695 *blank_pgdir = 1;
696 }
697#else
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700698 *blank_pgdir = 1;
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600699#endif
700 }
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700701 }
Rusty Russellbff672e2007-07-26 10:41:04 -0700702 /* Record which Guest toplevel this shadows. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200703 cpu->lg->pgdirs[next].gpgdir = gpgdir;
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700704 /* Release all the non-kernel mappings. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200705 flush_user_mappings(cpu->lg, next);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700706
707 return next;
708}
709
Rusty Russell2e04ef72009-07-30 16:03:45 -0600710/*H:430
711 * (iv) Switching page tables
Rusty Russellbff672e2007-07-26 10:41:04 -0700712 *
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600713 * Now we've seen all the page table setting and manipulation, let's see
Rusty Russelle1e72962007-10-25 15:02:50 +1000714 * what happens when the Guest changes page tables (ie. changes the top-level
Rusty Russell2e04ef72009-07-30 16:03:45 -0600715 * pgdir). This occurs on almost every context switch.
716 */
Glauber de Oliveira Costa4665ac82008-01-07 11:05:35 -0200717void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700718{
719 int newpgdir, repin = 0;
720
Rusty Russellbff672e2007-07-26 10:41:04 -0700721 /* Look to see if we have this one already. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200722 newpgdir = find_pgdir(cpu->lg, pgtable);
Rusty Russell2e04ef72009-07-30 16:03:45 -0600723 /*
724 * If not, we allocate or mug an existing one: if it's a fresh one,
725 * repin gets set to 1.
726 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200727 if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs))
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200728 newpgdir = new_pgdir(cpu, pgtable, &repin);
Rusty Russellbff672e2007-07-26 10:41:04 -0700729 /* Change the current pgd index to the new one. */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -0200730 cpu->cpu_pgd = newpgdir;
Rusty Russellbff672e2007-07-26 10:41:04 -0700731 /* If it was completely blank, we map in the Guest kernel stack */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700732 if (repin)
Glauber de Oliveira Costa4665ac82008-01-07 11:05:35 -0200733 pin_stack_pages(cpu);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700734}
735
Rusty Russell2e04ef72009-07-30 16:03:45 -0600736/*H:470
737 * Finally, a routine which throws away everything: all PGD entries in all
Rusty Russelle1e72962007-10-25 15:02:50 +1000738 * the shadow page tables, including the Guest's kernel mappings. This is used
Rusty Russell2e04ef72009-07-30 16:03:45 -0600739 * when we destroy the Guest.
740 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700741static void release_all_pagetables(struct lguest *lg)
742{
743 unsigned int i, j;
744
Rusty Russellbff672e2007-07-26 10:41:04 -0700745 /* Every shadow pagetable this Guest has */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700746 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600747 if (lg->pgdirs[i].pgdir) {
748#ifdef CONFIG_X86_PAE
749 pgd_t *spgd;
750 pmd_t *pmdpage;
751 unsigned int k;
752
753 /* Get the last pmd page. */
754 spgd = lg->pgdirs[i].pgdir + SWITCHER_PGD_INDEX;
755 pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
756
Rusty Russell2e04ef72009-07-30 16:03:45 -0600757 /*
758 * And release the pmd entries of that pmd page,
759 * except for the switcher pmd.
760 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600761 for (k = 0; k < SWITCHER_PMD_INDEX; k++)
762 release_pmd(&pmdpage[k]);
763#endif
Rusty Russellbff672e2007-07-26 10:41:04 -0700764 /* Every PGD entry except the Switcher at the top */
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700765 for (j = 0; j < SWITCHER_PGD_INDEX; j++)
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600766 release_pgd(lg->pgdirs[i].pgdir + j);
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600767 }
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700768}
769
Rusty Russell2e04ef72009-07-30 16:03:45 -0600770/*
771 * We also throw away everything when a Guest tells us it's changed a kernel
Rusty Russellbff672e2007-07-26 10:41:04 -0700772 * mapping. Since kernel mappings are in every page table, it's easiest to
Rusty Russelle1e72962007-10-25 15:02:50 +1000773 * throw them all away. This traps the Guest in amber for a while as
Rusty Russell2e04ef72009-07-30 16:03:45 -0600774 * everything faults back in, but it's rare.
775 */
Glauber de Oliveira Costa4665ac82008-01-07 11:05:35 -0200776void guest_pagetable_clear_all(struct lg_cpu *cpu)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700777{
Glauber de Oliveira Costa4665ac82008-01-07 11:05:35 -0200778 release_all_pagetables(cpu->lg);
Rusty Russellbff672e2007-07-26 10:41:04 -0700779 /* We need the Guest kernel stack mapped again. */
Glauber de Oliveira Costa4665ac82008-01-07 11:05:35 -0200780 pin_stack_pages(cpu);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700781}
Rusty Russelle1e72962007-10-25 15:02:50 +1000782/*:*/
Rusty Russell2e04ef72009-07-30 16:03:45 -0600783
784/*M:009
785 * Since we throw away all mappings when a kernel mapping changes, our
Rusty Russelle1e72962007-10-25 15:02:50 +1000786 * performance sucks for guests using highmem. In fact, a guest with
787 * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is
788 * usually slower than a Guest with less memory.
789 *
790 * This, of course, cannot be fixed. It would take some kind of... well, I
Rusty Russell2e04ef72009-07-30 16:03:45 -0600791 * don't know, but the term "puissant code-fu" comes to mind.
792:*/
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700793
Rusty Russell2e04ef72009-07-30 16:03:45 -0600794/*H:420
795 * This is the routine which actually sets the page table entry for then
Rusty Russellbff672e2007-07-26 10:41:04 -0700796 * "idx"'th shadow page table.
797 *
798 * Normally, we can just throw out the old entry and replace it with 0: if they
799 * use it demand_page() will put the new entry in. We need to do this anyway:
800 * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
801 * is read from, and _PAGE_DIRTY when it's written to.
802 *
803 * But Avi Kivity pointed out that most Operating Systems (Linux included) set
804 * these bits on PTEs immediately anyway. This is done to save the CPU from
805 * having to update them, but it helps us the same way: if they set
806 * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
807 * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
808 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200809static void do_set_pte(struct lg_cpu *cpu, int idx,
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000810 unsigned long vaddr, pte_t gpte)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700811{
Rusty Russelle1e72962007-10-25 15:02:50 +1000812 /* Look up the matching shadow page directory entry. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200813 pgd_t *spgd = spgd_addr(cpu, idx, vaddr);
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600814#ifdef CONFIG_X86_PAE
815 pmd_t *spmd;
816#endif
Rusty Russellbff672e2007-07-26 10:41:04 -0700817
818 /* If the top level isn't present, there's no entry to update. */
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +1000819 if (pgd_flags(*spgd) & _PAGE_PRESENT) {
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600820#ifdef CONFIG_X86_PAE
821 spmd = spmd_addr(cpu, *spgd, vaddr);
822 if (pmd_flags(*spmd) & _PAGE_PRESENT) {
823#endif
Rusty Russell2e04ef72009-07-30 16:03:45 -0600824 /* Otherwise, start by releasing the existing entry. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600825 pte_t *spte = spte_addr(cpu, *spgd, vaddr);
826 release_pte(*spte);
Rusty Russellbff672e2007-07-26 10:41:04 -0700827
Rusty Russell2e04ef72009-07-30 16:03:45 -0600828 /*
829 * If they're setting this entry as dirty or accessed,
830 * we might as well put that entry they've given us in
831 * now. This shaves 10% off a copy-on-write
832 * micro-benchmark.
833 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600834 if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) {
835 check_gpte(cpu, gpte);
836 native_set_pte(spte,
837 gpte_to_spte(cpu, gpte,
838 pte_flags(gpte) & _PAGE_DIRTY));
Rusty Russell2e04ef72009-07-30 16:03:45 -0600839 } else {
840 /*
841 * Otherwise kill it and we can demand_page()
842 * it in later.
843 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600844 native_set_pte(spte, __pte(0));
Rusty Russell2e04ef72009-07-30 16:03:45 -0600845 }
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600846#ifdef CONFIG_X86_PAE
847 }
848#endif
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700849 }
850}
851
Rusty Russell2e04ef72009-07-30 16:03:45 -0600852/*H:410
853 * Updating a PTE entry is a little trickier.
Rusty Russellbff672e2007-07-26 10:41:04 -0700854 *
855 * We keep track of several different page tables (the Guest uses one for each
856 * process, so it makes sense to cache at least a few). Each of these have
857 * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
858 * all processes. So when the page table above that address changes, we update
859 * all the page tables, not just the current one. This is rare.
860 *
Rusty Russella6bd8e12008-03-28 11:05:53 -0500861 * The benefit is that when we have to track a new page table, we can keep all
Rusty Russell2e04ef72009-07-30 16:03:45 -0600862 * the kernel mappings. This speeds up context switch immensely.
863 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200864void guest_set_pte(struct lg_cpu *cpu,
Rusty Russellee3db0f2007-10-22 11:03:34 +1000865 unsigned long gpgdir, unsigned long vaddr, pte_t gpte)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700866{
Rusty Russell2e04ef72009-07-30 16:03:45 -0600867 /*
868 * Kernel mappings must be changed on all top levels. Slow, but doesn't
869 * happen often.
870 */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200871 if (vaddr >= cpu->lg->kernel_address) {
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700872 unsigned int i;
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200873 for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++)
874 if (cpu->lg->pgdirs[i].pgdir)
875 do_set_pte(cpu, i, vaddr, gpte);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700876 } else {
Rusty Russellbff672e2007-07-26 10:41:04 -0700877 /* Is this page table one we have a shadow for? */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200878 int pgdir = find_pgdir(cpu->lg, gpgdir);
879 if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs))
Rusty Russellbff672e2007-07-26 10:41:04 -0700880 /* If so, do the update. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -0200881 do_set_pte(cpu, pgdir, vaddr, gpte);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700882 }
883}
884
Rusty Russellbff672e2007-07-26 10:41:04 -0700885/*H:400
Rusty Russelle1e72962007-10-25 15:02:50 +1000886 * (iii) Setting up a page table entry when the Guest tells us one has changed.
Rusty Russellbff672e2007-07-26 10:41:04 -0700887 *
888 * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
889 * with the other side of page tables while we're here: what happens when the
890 * Guest asks for a page table to be updated?
891 *
892 * We already saw that demand_page() will fill in the shadow page tables when
893 * needed, so we can simply remove shadow page table entries whenever the Guest
894 * tells us they've changed. When the Guest tries to use the new entry it will
895 * fault and demand_page() will fix it up.
896 *
897 * So with that in mind here's our code to to update a (top-level) PGD entry:
898 */
Matias Zabaljaureguiebe0ba82009-05-30 15:48:08 -0300899void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx)
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700900{
901 int pgdir;
902
903 if (idx >= SWITCHER_PGD_INDEX)
904 return;
905
Rusty Russellbff672e2007-07-26 10:41:04 -0700906 /* If they're talking about a page table we have a shadow for... */
Rusty Russellee3db0f2007-10-22 11:03:34 +1000907 pgdir = find_pgdir(lg, gpgdir);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700908 if (pgdir < ARRAY_SIZE(lg->pgdirs))
Rusty Russellbff672e2007-07-26 10:41:04 -0700909 /* ... throw it away. */
Matias Zabaljauregui90603d12009-06-12 22:27:06 -0600910 release_pgd(lg->pgdirs[pgdir].pgdir + idx);
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700911}
Rusty Russella91d74a2009-07-30 16:03:45 -0600912
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600913#ifdef CONFIG_X86_PAE
Rusty Russella91d74a2009-07-30 16:03:45 -0600914/* For setting a mid-level, we just throw everything away. It's easy. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600915void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx)
916{
917 guest_pagetable_clear_all(&lg->cpus[0]);
918}
919#endif
Rusty Russelld7e28ff2007-07-19 01:49:23 -0700920
Rusty Russella91d74a2009-07-30 16:03:45 -0600921/*H:505
922 * To get through boot, we construct simple identity page mappings (which
Rusty Russell2e04ef72009-07-30 16:03:45 -0600923 * set virtual == physical) and linear mappings which will get the Guest far
Rusty Russella91d74a2009-07-30 16:03:45 -0600924 * enough into the boot to create its own. The linear mapping means we
925 * simplify the Guest boot, but it makes assumptions about their PAGE_OFFSET,
926 * as you'll see.
Matias Zabaljauregui58a24562008-09-29 01:40:07 -0300927 *
928 * We lay them out of the way, just below the initrd (which is why we need to
Rusty Russell2e04ef72009-07-30 16:03:45 -0600929 * know its size here).
930 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -0300931static unsigned long setup_pagetables(struct lguest *lg,
932 unsigned long mem,
933 unsigned long initrd_size)
934{
935 pgd_t __user *pgdir;
936 pte_t __user *linear;
Matias Zabaljauregui58a24562008-09-29 01:40:07 -0300937 unsigned long mem_base = (unsigned long)lg->mem_base;
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600938 unsigned int mapped_pages, i, linear_pages;
939#ifdef CONFIG_X86_PAE
940 pmd_t __user *pmds;
941 unsigned int j;
942 pgd_t pgd;
943 pmd_t pmd;
944#else
945 unsigned int phys_linear;
946#endif
Matias Zabaljauregui58a24562008-09-29 01:40:07 -0300947
Rusty Russell2e04ef72009-07-30 16:03:45 -0600948 /*
949 * We have mapped_pages frames to map, so we need linear_pages page
950 * tables to map them.
951 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -0300952 mapped_pages = mem / PAGE_SIZE;
953 linear_pages = (mapped_pages + PTRS_PER_PTE - 1) / PTRS_PER_PTE;
954
955 /* We put the toplevel page directory page at the top of memory. */
956 pgdir = (pgd_t *)(mem + mem_base - initrd_size - PAGE_SIZE);
957
958 /* Now we use the next linear_pages pages as pte pages */
959 linear = (void *)pgdir - linear_pages * PAGE_SIZE;
960
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600961#ifdef CONFIG_X86_PAE
Rusty Russella91d74a2009-07-30 16:03:45 -0600962 /*
963 * And the single mid page goes below that. We only use one, but
964 * that's enough to map 1G, which definitely gets us through boot.
965 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600966 pmds = (void *)linear - PAGE_SIZE;
967#endif
Rusty Russell2e04ef72009-07-30 16:03:45 -0600968 /*
969 * Linear mapping is easy: put every page's address into the
970 * mapping in order.
971 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -0300972 for (i = 0; i < mapped_pages; i++) {
973 pte_t pte;
974 pte = pfn_pte(i, __pgprot(_PAGE_PRESENT|_PAGE_RW|_PAGE_USER));
975 if (copy_to_user(&linear[i], &pte, sizeof(pte)) != 0)
976 return -EFAULT;
977 }
978
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600979#ifdef CONFIG_X86_PAE
Rusty Russella91d74a2009-07-30 16:03:45 -0600980 /*
981 * Make the Guest PMD entries point to the corresponding place in the
982 * linear mapping (up to one page worth of PMD).
983 */
Rusty Russell92b4d8d2009-06-12 22:27:08 -0600984 for (i = j = 0; i < mapped_pages && j < PTRS_PER_PMD;
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600985 i += PTRS_PER_PTE, j++) {
Rusty Russella91d74a2009-07-30 16:03:45 -0600986 /* FIXME: native_set_pmd is overkill here. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600987 native_set_pmd(&pmd, __pmd(((unsigned long)(linear + i)
988 - mem_base) | _PAGE_PRESENT | _PAGE_RW | _PAGE_USER));
989
990 if (copy_to_user(&pmds[j], &pmd, sizeof(pmd)) != 0)
991 return -EFAULT;
992 }
993
Rusty Russella91d74a2009-07-30 16:03:45 -0600994 /* One PGD entry, pointing to that PMD page. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600995 set_pgd(&pgd, __pgd(((u32)pmds - mem_base) | _PAGE_PRESENT));
Rusty Russella91d74a2009-07-30 16:03:45 -0600996 /* Copy it in as the first PGD entry (ie. addresses 0-1G). */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -0600997 if (copy_to_user(&pgdir[0], &pgd, sizeof(pgd)) != 0)
998 return -EFAULT;
Rusty Russella91d74a2009-07-30 16:03:45 -0600999 /*
1000 * And the third PGD entry (ie. addresses 3G-4G).
1001 *
1002 * FIXME: This assumes that PAGE_OFFSET for the Guest is 0xC0000000.
1003 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001004 if (copy_to_user(&pgdir[3], &pgd, sizeof(pgd)) != 0)
1005 return -EFAULT;
1006#else
Rusty Russella91d74a2009-07-30 16:03:45 -06001007 /*
1008 * The top level points to the linear page table pages above.
1009 * We setup the identity and linear mappings here.
1010 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001011 phys_linear = (unsigned long)linear - mem_base;
1012 for (i = 0; i < mapped_pages; i += PTRS_PER_PTE) {
1013 pgd_t pgd;
Rusty Russella91d74a2009-07-30 16:03:45 -06001014 /*
1015 * Create a PGD entry which points to the right part of the
1016 * linear PTE pages.
1017 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001018 pgd = __pgd((phys_linear + i * sizeof(pte_t)) |
1019 (_PAGE_PRESENT | _PAGE_RW | _PAGE_USER));
1020
Rusty Russella91d74a2009-07-30 16:03:45 -06001021 /*
1022 * Copy it into the PGD page at 0 and PAGE_OFFSET.
1023 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001024 if (copy_to_user(&pgdir[i / PTRS_PER_PTE], &pgd, sizeof(pgd))
1025 || copy_to_user(&pgdir[pgd_index(PAGE_OFFSET)
1026 + i / PTRS_PER_PTE],
1027 &pgd, sizeof(pgd)))
1028 return -EFAULT;
1029 }
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001030#endif
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001031
Rusty Russell2e04ef72009-07-30 16:03:45 -06001032 /*
Rusty Russella91d74a2009-07-30 16:03:45 -06001033 * We return the top level (guest-physical) address: we remember where
1034 * this is to write it into lguest_data when the Guest initializes.
Rusty Russell2e04ef72009-07-30 16:03:45 -06001035 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001036 return (unsigned long)pgdir - mem_base;
1037}
1038
Rusty Russell2e04ef72009-07-30 16:03:45 -06001039/*H:500
1040 * (vii) Setting up the page tables initially.
Rusty Russellbff672e2007-07-26 10:41:04 -07001041 *
1042 * When a Guest is first created, the Launcher tells us where the toplevel of
Rusty Russell2e04ef72009-07-30 16:03:45 -06001043 * its first page table is. We set some things up here:
1044 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001045int init_guest_pagetable(struct lguest *lg)
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001046{
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001047 u64 mem;
1048 u32 initrd_size;
1049 struct boot_params __user *boot = (struct boot_params *)lg->mem_base;
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001050#ifdef CONFIG_X86_PAE
1051 pgd_t *pgd;
1052 pmd_t *pmd_table;
1053#endif
Rusty Russell2e04ef72009-07-30 16:03:45 -06001054 /*
1055 * Get the Guest memory size and the ramdisk size from the boot header
1056 * located at lg->mem_base (Guest address 0).
1057 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001058 if (copy_from_user(&mem, &boot->e820_map[0].size, sizeof(mem))
1059 || get_user(initrd_size, &boot->hdr.ramdisk_size))
1060 return -EFAULT;
1061
Rusty Russell2e04ef72009-07-30 16:03:45 -06001062 /*
1063 * We start on the first shadow page table, and give it a blank PGD
1064 * page.
1065 */
Matias Zabaljauregui58a24562008-09-29 01:40:07 -03001066 lg->pgdirs[0].gpgdir = setup_pagetables(lg, mem, initrd_size);
1067 if (IS_ERR_VALUE(lg->pgdirs[0].gpgdir))
1068 return lg->pgdirs[0].gpgdir;
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -02001069 lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL);
1070 if (!lg->pgdirs[0].pgdir)
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001071 return -ENOMEM;
Rusty Russella91d74a2009-07-30 16:03:45 -06001072
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001073#ifdef CONFIG_X86_PAE
Rusty Russella91d74a2009-07-30 16:03:45 -06001074 /* For PAE, we also create the initial mid-level. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001075 pgd = lg->pgdirs[0].pgdir;
1076 pmd_table = (pmd_t *) get_zeroed_page(GFP_KERNEL);
1077 if (!pmd_table)
1078 return -ENOMEM;
1079
1080 set_pgd(pgd + SWITCHER_PGD_INDEX,
1081 __pgd(__pa(pmd_table) | _PAGE_PRESENT));
1082#endif
Rusty Russella91d74a2009-07-30 16:03:45 -06001083
1084 /* This is the current page table. */
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -02001085 lg->cpus[0].cpu_pgd = 0;
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001086 return 0;
1087}
1088
Rusty Russella91d74a2009-07-30 16:03:45 -06001089/*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -02001090void page_table_guest_data_init(struct lg_cpu *cpu)
Rusty Russell47436aa2007-10-22 11:03:36 +10001091{
1092 /* We get the kernel address: above this is all kernel memory. */
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -02001093 if (get_user(cpu->lg->kernel_address,
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001094 &cpu->lg->lguest_data->kernel_address)
Rusty Russell2e04ef72009-07-30 16:03:45 -06001095 /*
1096 * We tell the Guest that it can't use the top 2 or 4 MB
1097 * of virtual addresses used by the Switcher.
1098 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001099 || put_user(RESERVE_MEM * 1024 * 1024,
1100 &cpu->lg->lguest_data->reserve_mem)
1101 || put_user(cpu->lg->pgdirs[0].gpgdir,
1102 &cpu->lg->lguest_data->pgdir))
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -02001103 kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data);
Rusty Russell47436aa2007-10-22 11:03:36 +10001104
Rusty Russell2e04ef72009-07-30 16:03:45 -06001105 /*
1106 * In flush_user_mappings() we loop from 0 to
Rusty Russell47436aa2007-10-22 11:03:36 +10001107 * "pgd_index(lg->kernel_address)". This assumes it won't hit the
Rusty Russell2e04ef72009-07-30 16:03:45 -06001108 * Switcher mappings, so check that now.
1109 */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001110#ifdef CONFIG_X86_PAE
1111 if (pgd_index(cpu->lg->kernel_address) == SWITCHER_PGD_INDEX &&
1112 pmd_index(cpu->lg->kernel_address) == SWITCHER_PMD_INDEX)
1113#else
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -02001114 if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX)
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001115#endif
Glauber de Oliveira Costa382ac6b2008-01-17 19:19:42 -02001116 kill_guest(cpu, "bad kernel address %#lx",
1117 cpu->lg->kernel_address);
Rusty Russell47436aa2007-10-22 11:03:36 +10001118}
1119
Rusty Russellbff672e2007-07-26 10:41:04 -07001120/* When a Guest dies, our cleanup is fairly simple. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001121void free_guest_pagetable(struct lguest *lg)
1122{
1123 unsigned int i;
1124
Rusty Russellbff672e2007-07-26 10:41:04 -07001125 /* Throw away all page table pages. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001126 release_all_pagetables(lg);
Rusty Russellbff672e2007-07-26 10:41:04 -07001127 /* Now free the top levels: free_page() can handle 0 just fine. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001128 for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
1129 free_page((long)lg->pgdirs[i].pgdir);
1130}
1131
Rusty Russell2e04ef72009-07-30 16:03:45 -06001132/*H:480
1133 * (vi) Mapping the Switcher when the Guest is about to run.
Rusty Russellbff672e2007-07-26 10:41:04 -07001134 *
Rusty Russelle1e72962007-10-25 15:02:50 +10001135 * The Switcher and the two pages for this CPU need to be visible in the
Rusty Russellbff672e2007-07-26 10:41:04 -07001136 * Guest (and not the pages for other CPUs). We have the appropriate PTE pages
Rusty Russelle1e72962007-10-25 15:02:50 +10001137 * for each CPU already set up, we just need to hook them in now we know which
Rusty Russell2e04ef72009-07-30 16:03:45 -06001138 * Guest is about to run on this CPU.
1139 */
Glauber de Oliveira Costa0c784412008-01-07 11:05:30 -02001140void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages)
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001141{
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +10001142 pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages);
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +10001143 pte_t regs_pte;
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -02001144 unsigned long pfn;
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001145
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001146#ifdef CONFIG_X86_PAE
1147 pmd_t switcher_pmd;
1148 pmd_t *pmd_table;
1149
Rusty Russella91d74a2009-07-30 16:03:45 -06001150 /* FIXME: native_set_pmd is overkill here. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001151 native_set_pmd(&switcher_pmd, pfn_pmd(__pa(switcher_pte_page) >>
1152 PAGE_SHIFT, PAGE_KERNEL_EXEC));
1153
Rusty Russella91d74a2009-07-30 16:03:45 -06001154 /* Figure out where the pmd page is, by reading the PGD, and converting
1155 * it to a virtual address. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001156 pmd_table = __va(pgd_pfn(cpu->lg->
1157 pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX])
1158 << PAGE_SHIFT);
Rusty Russella91d74a2009-07-30 16:03:45 -06001159 /* Now write it into the shadow page table. */
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001160 native_set_pmd(&pmd_table[SWITCHER_PMD_INDEX], switcher_pmd);
1161#else
1162 pgd_t switcher_pgd;
1163
Rusty Russell2e04ef72009-07-30 16:03:45 -06001164 /*
1165 * Make the last PGD entry for this Guest point to the Switcher's PTE
1166 * page for this CPU (with appropriate flags).
1167 */
Matias Zabaljaureguied1dc772009-05-30 15:35:49 -03001168 switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL_EXEC);
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +10001169
Glauber de Oliveira Costa17136082008-01-07 11:05:37 -02001170 cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd;
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001171
Matias Zabaljaureguiacdd0b62009-06-12 22:27:07 -06001172#endif
Rusty Russell2e04ef72009-07-30 16:03:45 -06001173 /*
1174 * We also change the Switcher PTE page. When we're running the Guest,
Rusty Russellbff672e2007-07-26 10:41:04 -07001175 * we want the Guest's "regs" page to appear where the first Switcher
1176 * page for this CPU is. This is an optimization: when the Switcher
1177 * saves the Guest registers, it saves them into the first page of this
1178 * CPU's "struct lguest_pages": if we make sure the Guest's register
1179 * page is already mapped there, we don't have to copy them out
Rusty Russell2e04ef72009-07-30 16:03:45 -06001180 * again.
1181 */
Glauber de Oliveira Costaa53a35a2008-01-07 11:05:32 -02001182 pfn = __pa(cpu->regs_page) >> PAGE_SHIFT;
Matias Zabaljauregui90603d12009-06-12 22:27:06 -06001183 native_set_pte(&regs_pte, pfn_pte(pfn, PAGE_KERNEL));
1184 native_set_pte(&switcher_pte_page[pte_index((unsigned long)pages)],
1185 regs_pte);
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001186}
Rusty Russellbff672e2007-07-26 10:41:04 -07001187/*:*/
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001188
1189static void free_switcher_pte_pages(void)
1190{
1191 unsigned int i;
1192
1193 for_each_possible_cpu(i)
1194 free_page((long)switcher_pte_page(i));
1195}
1196
Rusty Russell2e04ef72009-07-30 16:03:45 -06001197/*H:520
1198 * Setting up the Switcher PTE page for given CPU is fairly easy, given
Rusty Russellbff672e2007-07-26 10:41:04 -07001199 * the CPU number and the "struct page"s for the Switcher code itself.
1200 *
Rusty Russell2e04ef72009-07-30 16:03:45 -06001201 * Currently the Switcher is less than a page long, so "pages" is always 1.
1202 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001203static __init void populate_switcher_pte_page(unsigned int cpu,
1204 struct page *switcher_page[],
1205 unsigned int pages)
1206{
1207 unsigned int i;
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +10001208 pte_t *pte = switcher_pte_page(cpu);
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001209
Rusty Russellbff672e2007-07-26 10:41:04 -07001210 /* The first entries are easy: they map the Switcher code. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001211 for (i = 0; i < pages; i++) {
Matias Zabaljauregui90603d12009-06-12 22:27:06 -06001212 native_set_pte(&pte[i], mk_pte(switcher_page[i],
1213 __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)));
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001214 }
1215
Rusty Russellbff672e2007-07-26 10:41:04 -07001216 /* The only other thing we map is this CPU's pair of pages. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001217 i = pages + cpu*2;
1218
Rusty Russellbff672e2007-07-26 10:41:04 -07001219 /* First page (Guest registers) is writable from the Guest */
Matias Zabaljauregui90603d12009-06-12 22:27:06 -06001220 native_set_pte(&pte[i], pfn_pte(page_to_pfn(switcher_page[i]),
1221 __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW)));
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +10001222
Rusty Russell2e04ef72009-07-30 16:03:45 -06001223 /*
1224 * The second page contains the "struct lguest_ro_state", and is
1225 * read-only.
1226 */
Matias Zabaljauregui90603d12009-06-12 22:27:06 -06001227 native_set_pte(&pte[i+1], pfn_pte(page_to_pfn(switcher_page[i+1]),
1228 __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)));
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001229}
1230
Rusty Russell2e04ef72009-07-30 16:03:45 -06001231/*
1232 * We've made it through the page table code. Perhaps our tired brains are
Rusty Russelle1e72962007-10-25 15:02:50 +10001233 * still processing the details, or perhaps we're simply glad it's over.
1234 *
Rusty Russella6bd8e12008-03-28 11:05:53 -05001235 * If nothing else, note that all this complexity in juggling shadow page tables
1236 * in sync with the Guest's page tables is for one reason: for most Guests this
1237 * page table dance determines how bad performance will be. This is why Xen
1238 * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD
1239 * have implemented shadow page table support directly into hardware.
Rusty Russelle1e72962007-10-25 15:02:50 +10001240 *
Rusty Russell2e04ef72009-07-30 16:03:45 -06001241 * There is just one file remaining in the Host.
1242 */
Rusty Russelle1e72962007-10-25 15:02:50 +10001243
Rusty Russell2e04ef72009-07-30 16:03:45 -06001244/*H:510
1245 * At boot or module load time, init_pagetables() allocates and populates
1246 * the Switcher PTE page for each CPU.
1247 */
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001248__init int init_pagetables(struct page **switcher_page, unsigned int pages)
1249{
1250 unsigned int i;
1251
1252 for_each_possible_cpu(i) {
Matias Zabaljaureguidf29f432007-10-22 11:03:33 +10001253 switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL);
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001254 if (!switcher_pte_page(i)) {
1255 free_switcher_pte_pages();
1256 return -ENOMEM;
1257 }
1258 populate_switcher_pte_page(i, switcher_page, pages);
1259 }
1260 return 0;
1261}
Rusty Russellbff672e2007-07-26 10:41:04 -07001262/*:*/
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001263
Rusty Russellbff672e2007-07-26 10:41:04 -07001264/* Cleaning up simply involves freeing the PTE page for each CPU. */
Rusty Russelld7e28ff2007-07-19 01:49:23 -07001265void free_pagetables(void)
1266{
1267 free_switcher_pte_pages();
1268}