Rusty Russell | f938d2c | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 1 | /*P:700 The pagetable code, on the other hand, still shows the scars of |
| 2 | * previous encounters. It's functional, and as neat as it can be in the |
| 3 | * circumstances, but be wary, for these things are subtle and break easily. |
| 4 | * The Guest provides a virtual to physical mapping, but we can neither trust |
| 5 | * it nor use it: we verify and convert it here to point the hardware to the |
| 6 | * actual Guest pages when running the Guest. :*/ |
| 7 | |
| 8 | /* Copyright (C) Rusty Russell IBM Corporation 2006. |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 9 | * GPL v2 and any later version */ |
| 10 | #include <linux/mm.h> |
| 11 | #include <linux/types.h> |
| 12 | #include <linux/spinlock.h> |
| 13 | #include <linux/random.h> |
| 14 | #include <linux/percpu.h> |
| 15 | #include <asm/tlbflush.h> |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 16 | #include <asm/uaccess.h> |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 17 | #include "lg.h" |
| 18 | |
Rusty Russell | f56a384 | 2007-07-26 10:41:05 -0700 | [diff] [blame] | 19 | /*M:008 We hold reference to pages, which prevents them from being swapped. |
| 20 | * It'd be nice to have a callback in the "struct mm_struct" when Linux wants |
| 21 | * to swap out. If we had this, and a shrinker callback to trim PTE pages, we |
| 22 | * could probably consider launching Guests as non-root. :*/ |
| 23 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 24 | /*H:300 |
| 25 | * The Page Table Code |
| 26 | * |
| 27 | * We use two-level page tables for the Guest. If you're not entirely |
| 28 | * comfortable with virtual addresses, physical addresses and page tables then |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 29 | * I recommend you review arch/x86/lguest/boot.c's "Page Table Handling" (with |
| 30 | * diagrams!). |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 31 | * |
| 32 | * The Guest keeps page tables, but we maintain the actual ones here: these are |
| 33 | * called "shadow" page tables. Which is a very Guest-centric name: these are |
| 34 | * the real page tables the CPU uses, although we keep them up to date to |
| 35 | * reflect the Guest's. (See what I mean about weird naming? Since when do |
| 36 | * shadows reflect anything?) |
| 37 | * |
| 38 | * Anyway, this is the most complicated part of the Host code. There are seven |
| 39 | * parts to this: |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 40 | * (i) Looking up a page table entry when the Guest faults, |
| 41 | * (ii) Making sure the Guest stack is mapped, |
| 42 | * (iii) Setting up a page table entry when the Guest tells us one has changed, |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 43 | * (iv) Switching page tables, |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 44 | * (v) Flushing (throwing away) page tables, |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 45 | * (vi) Mapping the Switcher when the Guest is about to run, |
| 46 | * (vii) Setting up the page tables initially. |
| 47 | :*/ |
| 48 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 49 | |
| 50 | /* 1024 entries in a page table page maps 1024 pages: 4MB. The Switcher is |
| 51 | * conveniently placed at the top 4MB, so it uses a separate, complete PTE |
| 52 | * page. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 53 | #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 54 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 55 | /* We actually need a separate PTE page for each CPU. Remember that after the |
| 56 | * Switcher code itself comes two pages for each CPU, and we don't want this |
| 57 | * CPU's guest to see the pages of any other CPU. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 58 | static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 59 | #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) |
| 60 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 61 | /*H:320 The page table code is curly enough to need helper functions to keep it |
| 62 | * clear and clean. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 63 | * |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 64 | * There are two functions which return pointers to the shadow (aka "real") |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 65 | * page tables. |
| 66 | * |
| 67 | * spgd_addr() takes the virtual address and returns a pointer to the top-level |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 68 | * page directory entry (PGD) for that address. Since we keep track of several |
| 69 | * page tables, the "i" argument tells us which one we're interested in (it's |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 70 | * usually the current one). */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 71 | static pgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 72 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 73 | unsigned int index = pgd_index(vaddr); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 74 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 75 | /* We kill any Guest trying to touch the Switcher addresses. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 76 | if (index >= SWITCHER_PGD_INDEX) { |
| 77 | kill_guest(lg, "attempt to access switcher pages"); |
| 78 | index = 0; |
| 79 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 80 | /* Return a pointer index'th pgd entry for the i'th page table. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 81 | return &lg->pgdirs[i].pgdir[index]; |
| 82 | } |
| 83 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 84 | /* This routine then takes the page directory entry returned above, which |
| 85 | * contains the address of the page table entry (PTE) page. It then returns a |
| 86 | * pointer to the PTE entry for the given address. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 87 | static pte_t *spte_addr(struct lguest *lg, pgd_t spgd, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 88 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 89 | pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 90 | /* You should never call this if the PGD entry wasn't valid */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 91 | BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); |
| 92 | return &page[(vaddr >> PAGE_SHIFT) % PTRS_PER_PTE]; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 93 | } |
| 94 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 95 | /* These two functions just like the above two, except they access the Guest |
| 96 | * page tables. Hence they return a Guest address. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 97 | static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 98 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 99 | unsigned int index = vaddr >> (PGDIR_SHIFT); |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 100 | return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 101 | } |
| 102 | |
| 103 | static unsigned long gpte_addr(struct lguest *lg, |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 104 | pgd_t gpgd, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 105 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 106 | unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; |
| 107 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); |
| 108 | return gpage + ((vaddr>>PAGE_SHIFT) % PTRS_PER_PTE) * sizeof(pte_t); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 109 | } |
| 110 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 111 | /*H:350 This routine takes a page number given by the Guest and converts it to |
| 112 | * an actual, physical page number. It can fail for several reasons: the |
| 113 | * virtual address might not be mapped by the Launcher, the write flag is set |
| 114 | * and the page is read-only, or the write flag was set and the page was |
| 115 | * shared so had to be copied, but we ran out of memory. |
| 116 | * |
| 117 | * This holds a reference to the page, so release_pte() is careful to |
| 118 | * put that back. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 119 | static unsigned long get_pfn(unsigned long virtpfn, int write) |
| 120 | { |
| 121 | struct page *page; |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 122 | /* This value indicates failure. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 123 | unsigned long ret = -1UL; |
| 124 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 125 | /* get_user_pages() is a complex interface: it gets the "struct |
| 126 | * vm_area_struct" and "struct page" assocated with a range of pages. |
| 127 | * It also needs the task's mmap_sem held, and is not very quick. |
| 128 | * It returns the number of pages it got. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 129 | down_read(¤t->mm->mmap_sem); |
| 130 | if (get_user_pages(current, current->mm, virtpfn << PAGE_SHIFT, |
| 131 | 1, write, 1, &page, NULL) == 1) |
| 132 | ret = page_to_pfn(page); |
| 133 | up_read(¤t->mm->mmap_sem); |
| 134 | return ret; |
| 135 | } |
| 136 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 137 | /*H:340 Converting a Guest page table entry to a shadow (ie. real) page table |
| 138 | * entry can be a little tricky. The flags are (almost) the same, but the |
| 139 | * Guest PTE contains a virtual page number: the CPU needs the real page |
| 140 | * number. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 141 | static pte_t gpte_to_spte(struct lguest *lg, pte_t gpte, int write) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 142 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 143 | unsigned long pfn, base, flags; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 144 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 145 | /* The Guest sets the global flag, because it thinks that it is using |
| 146 | * PGE. We only told it to use PGE so it would tell us whether it was |
| 147 | * flushing a kernel mapping or a userspace mapping. We don't actually |
| 148 | * use the global bit, so throw it away. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 149 | flags = (pte_flags(gpte) & ~_PAGE_GLOBAL); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 150 | |
Rusty Russell | 3c6b5bf | 2007-10-22 11:03:26 +1000 | [diff] [blame] | 151 | /* The Guest's pages are offset inside the Launcher. */ |
| 152 | base = (unsigned long)lg->mem_base / PAGE_SIZE; |
| 153 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 154 | /* We need a temporary "unsigned long" variable to hold the answer from |
| 155 | * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't |
| 156 | * fit in spte.pfn. get_pfn() finds the real physical number of the |
| 157 | * page, given the virtual number. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 158 | pfn = get_pfn(base + pte_pfn(gpte), write); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 159 | if (pfn == -1UL) { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 160 | kill_guest(lg, "failed to get page %lu", pte_pfn(gpte)); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 161 | /* When we destroy the Guest, we'll go through the shadow page |
| 162 | * tables and release_pte() them. Make sure we don't think |
| 163 | * this one is valid! */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 164 | flags = 0; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 165 | } |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 166 | /* Now we assemble our shadow PTE from the page number and flags. */ |
| 167 | return pfn_pte(pfn, __pgprot(flags)); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 168 | } |
| 169 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 170 | /*H:460 And to complete the chain, release_pte() looks like this: */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 171 | static void release_pte(pte_t pte) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 172 | { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 173 | /* Remember that get_user_pages() took a reference to the page, in |
| 174 | * get_pfn()? We have to put it back now. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 175 | if (pte_flags(pte) & _PAGE_PRESENT) |
| 176 | put_page(pfn_to_page(pte_pfn(pte))); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 177 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 178 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 179 | |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 180 | static void check_gpte(struct lguest *lg, pte_t gpte) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 181 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 182 | if ((pte_flags(gpte) & (_PAGE_PWT|_PAGE_PSE)) |
| 183 | || pte_pfn(gpte) >= lg->pfn_limit) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 184 | kill_guest(lg, "bad page table entry"); |
| 185 | } |
| 186 | |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 187 | static void check_gpgd(struct lguest *lg, pgd_t gpgd) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 188 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 189 | if ((pgd_flags(gpgd) & ~_PAGE_TABLE) || pgd_pfn(gpgd) >= lg->pfn_limit) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 190 | kill_guest(lg, "bad page directory entry"); |
| 191 | } |
| 192 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 193 | /*H:330 |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 194 | * (i) Looking up a page table entry when the Guest faults. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 195 | * |
| 196 | * We saw this call in run_guest(): when we see a page fault in the Guest, we |
| 197 | * come here. That's because we only set up the shadow page tables lazily as |
| 198 | * they're needed, so we get page faults all the time and quietly fix them up |
| 199 | * and return to the Guest without it knowing. |
| 200 | * |
| 201 | * If we fixed up the fault (ie. we mapped the address), this routine returns |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 202 | * true. Otherwise, it was a real fault and we need to tell the Guest. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 203 | int demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 204 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 205 | pgd_t gpgd; |
| 206 | pgd_t *spgd; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 207 | unsigned long gpte_ptr; |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 208 | pte_t gpte; |
| 209 | pte_t *spte; |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 210 | struct lguest *lg = cpu->lg; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 211 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 212 | /* First step: get the top-level Guest page table entry. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 213 | gpgd = lgread(lg, gpgd_addr(cpu, vaddr), pgd_t); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 214 | /* Toplevel not present? We can't map it in. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 215 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 216 | return 0; |
| 217 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 218 | /* Now look at the matching shadow entry. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 219 | spgd = spgd_addr(lg, cpu->cpu_pgd, vaddr); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 220 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 221 | /* No shadow entry: allocate a new shadow PTE page. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 222 | unsigned long ptepage = get_zeroed_page(GFP_KERNEL); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 223 | /* This is not really the Guest's fault, but killing it is |
| 224 | * simple for this corner case. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 225 | if (!ptepage) { |
| 226 | kill_guest(lg, "out of memory allocating pte page"); |
| 227 | return 0; |
| 228 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 229 | /* We check that the Guest pgd is OK. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 230 | check_gpgd(lg, gpgd); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 231 | /* And we copy the flags to the shadow PGD entry. The page |
| 232 | * number in the shadow PGD is the page we just allocated. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 233 | *spgd = __pgd(__pa(ptepage) | pgd_flags(gpgd)); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 234 | } |
| 235 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 236 | /* OK, now we look at the lower level in the Guest page table: keep its |
| 237 | * address, because we might update it later. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 238 | gpte_ptr = gpte_addr(lg, gpgd, vaddr); |
Rusty Russell | 2d37f94 | 2007-10-22 11:24:24 +1000 | [diff] [blame] | 239 | gpte = lgread(lg, gpte_ptr, pte_t); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 240 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 241 | /* If this page isn't in the Guest page tables, we can't page it in. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 242 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 243 | return 0; |
| 244 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 245 | /* Check they're not trying to write to a page the Guest wants |
| 246 | * read-only (bit 2 of errcode == write). */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 247 | if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW)) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 248 | return 0; |
| 249 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 250 | /* User access to a kernel-only page? (bit 3 == user access) */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 251 | if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER)) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 252 | return 0; |
| 253 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 254 | /* Check that the Guest PTE flags are OK, and the page number is below |
| 255 | * the pfn_limit (ie. not mapping the Launcher binary). */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 256 | check_gpte(lg, gpte); |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 257 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 258 | /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 259 | gpte = pte_mkyoung(gpte); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 260 | if (errcode & 2) |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 261 | gpte = pte_mkdirty(gpte); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 262 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 263 | /* Get the pointer to the shadow PTE entry we're going to set. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 264 | spte = spte_addr(lg, *spgd, vaddr); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 265 | /* If there was a valid shadow PTE entry here before, we release it. |
| 266 | * This can happen with a write to a previously read-only entry. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 267 | release_pte(*spte); |
| 268 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 269 | /* If this is a write, we insist that the Guest page is writable (the |
| 270 | * final arg to gpte_to_spte()). */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 271 | if (pte_dirty(gpte)) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 272 | *spte = gpte_to_spte(lg, gpte, 1); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 273 | else |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 274 | /* If this is a read, don't set the "writable" bit in the page |
| 275 | * table entry, even if the Guest says it's writable. That way |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 276 | * we will come back here when a write does actually occur, so |
| 277 | * we can update the Guest's _PAGE_DIRTY flag. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 278 | *spte = gpte_to_spte(lg, pte_wrprotect(gpte), 0); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 279 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 280 | /* Finally, we write the Guest PTE entry back: we've set the |
| 281 | * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */ |
Rusty Russell | 2d37f94 | 2007-10-22 11:24:24 +1000 | [diff] [blame] | 282 | lgwrite(lg, gpte_ptr, pte_t, gpte); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 283 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 284 | /* The fault is fixed, the page table is populated, the mapping |
| 285 | * manipulated, the result returned and the code complete. A small |
| 286 | * delay and a trace of alliteration are the only indications the Guest |
| 287 | * has that a page fault occurred at all. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 288 | return 1; |
| 289 | } |
| 290 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 291 | /*H:360 |
| 292 | * (ii) Making sure the Guest stack is mapped. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 293 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 294 | * Remember that direct traps into the Guest need a mapped Guest kernel stack. |
| 295 | * pin_stack_pages() calls us here: we could simply call demand_page(), but as |
| 296 | * we've seen that logic is quite long, and usually the stack pages are already |
| 297 | * mapped, so it's overkill. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 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? */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 301 | static int page_writable(struct lg_cpu *cpu, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 302 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 303 | pgd_t *spgd; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 304 | unsigned long flags; |
| 305 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 306 | /* Look at the current top level entry: is it present? */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 307 | spgd = spgd_addr(cpu->lg, cpu->cpu_pgd, vaddr); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 308 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 309 | return 0; |
| 310 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 311 | /* Check the flags on the pte entry itself: it must be present and |
| 312 | * writable. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 313 | flags = pte_flags(*(spte_addr(cpu->lg, *spgd, vaddr))); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 314 | |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 315 | return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); |
| 316 | } |
| 317 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 318 | /* So, when pin_stack_pages() asks us to pin a page, we check if it's already |
| 319 | * in the page tables, and if not, we call demand_page() with error code 2 |
| 320 | * (meaning "write"). */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 321 | void pin_page(struct lg_cpu *cpu, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 322 | { |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 323 | if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2)) |
| 324 | kill_guest(cpu->lg, "bad stack page %#lx", vaddr); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 325 | } |
| 326 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 327 | /*H:450 If we chase down the release_pgd() code, it looks like this: */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 328 | static void release_pgd(struct lguest *lg, pgd_t *spgd) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 329 | { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 330 | /* If the entry's not present, there's nothing to release. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 331 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 332 | unsigned int i; |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 333 | /* Converting the pfn to find the actual PTE page is easy: turn |
| 334 | * the page number into a physical address, then convert to a |
| 335 | * virtual address (easy for kernel pages like this one). */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 336 | pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 337 | /* For each entry in the page, we might need to release it. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 338 | for (i = 0; i < PTRS_PER_PTE; i++) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 339 | release_pte(ptepage[i]); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 340 | /* Now we can free the page of PTEs */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 341 | free_page((long)ptepage); |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 342 | /* And zero out the PGD entry so we never release it twice. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 343 | *spgd = __pgd(0); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 344 | } |
| 345 | } |
| 346 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 347 | /*H:445 We saw flush_user_mappings() twice: once from the flush_user_mappings() |
| 348 | * hypercall and once in new_pgdir() when we re-used a top-level pgdir page. |
| 349 | * It simply releases every PTE page from 0 up to the Guest's kernel address. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 350 | static void flush_user_mappings(struct lguest *lg, int idx) |
| 351 | { |
| 352 | unsigned int i; |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 353 | /* Release every pgd entry up to the kernel's address. */ |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 354 | for (i = 0; i < pgd_index(lg->kernel_address); i++) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 355 | release_pgd(lg, lg->pgdirs[idx].pgdir + i); |
| 356 | } |
| 357 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 358 | /*H:440 (v) Flushing (throwing away) page tables, |
| 359 | * |
| 360 | * The Guest has a hypercall to throw away the page tables: it's used when a |
| 361 | * large number of mappings have been changed. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 362 | void guest_pagetable_flush_user(struct lg_cpu *cpu) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 363 | { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 364 | /* Drop the userspace part of the current page table. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 365 | flush_user_mappings(cpu->lg, cpu->cpu_pgd); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 366 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 367 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 368 | |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 369 | /* We walk down the guest page tables to get a guest-physical address */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 370 | unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr) |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 371 | { |
| 372 | pgd_t gpgd; |
| 373 | pte_t gpte; |
| 374 | |
| 375 | /* First step: get the top-level Guest page table entry. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 376 | gpgd = lgread(cpu->lg, gpgd_addr(cpu, vaddr), pgd_t); |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 377 | /* Toplevel not present? We can't map it in. */ |
| 378 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 379 | kill_guest(cpu->lg, "Bad address %#lx", vaddr); |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 380 | |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 381 | gpte = lgread(cpu->lg, gpte_addr(cpu->lg, gpgd, vaddr), pte_t); |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 382 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 383 | kill_guest(cpu->lg, "Bad address %#lx", vaddr); |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 384 | |
| 385 | return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK); |
| 386 | } |
| 387 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 388 | /* We keep several page tables. This is a simple routine to find the page |
| 389 | * table (if any) corresponding to this top-level address the Guest has given |
| 390 | * us. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 391 | static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) |
| 392 | { |
| 393 | unsigned int i; |
| 394 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 395 | if (lg->pgdirs[i].gpgdir == pgtable) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 396 | break; |
| 397 | return i; |
| 398 | } |
| 399 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 400 | /*H:435 And this is us, creating the new page directory. If we really do |
| 401 | * allocate a new one (and so the kernel parts are not there), we set |
| 402 | * blank_pgdir. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 403 | static unsigned int new_pgdir(struct lg_cpu *cpu, |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 404 | unsigned long gpgdir, |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 405 | int *blank_pgdir) |
| 406 | { |
| 407 | unsigned int next; |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 408 | struct lguest *lg = cpu->lg; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 409 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 410 | /* We pick one entry at random to throw out. Choosing the Least |
| 411 | * Recently Used might be better, but this is easy. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 412 | next = random32() % ARRAY_SIZE(lg->pgdirs); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 413 | /* If it's never been allocated at all before, try now. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 414 | if (!lg->pgdirs[next].pgdir) { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 415 | lg->pgdirs[next].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 416 | /* If the allocation fails, just keep using the one we have */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 417 | if (!lg->pgdirs[next].pgdir) |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 418 | next = cpu->cpu_pgd; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 419 | else |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 420 | /* This is a blank page, so there are no kernel |
| 421 | * mappings: caller must map the stack! */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 422 | *blank_pgdir = 1; |
| 423 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 424 | /* Record which Guest toplevel this shadows. */ |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 425 | lg->pgdirs[next].gpgdir = gpgdir; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 426 | /* Release all the non-kernel mappings. */ |
| 427 | flush_user_mappings(lg, next); |
| 428 | |
| 429 | return next; |
| 430 | } |
| 431 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 432 | /*H:430 (iv) Switching page tables |
| 433 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 434 | * Now we've seen all the page table setting and manipulation, let's see what |
| 435 | * what happens when the Guest changes page tables (ie. changes the top-level |
| 436 | * pgdir). This occurs on almost every context switch. */ |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 437 | void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 438 | { |
| 439 | int newpgdir, repin = 0; |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 440 | struct lguest *lg = cpu->lg; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 441 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 442 | /* Look to see if we have this one already. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 443 | newpgdir = find_pgdir(lg, pgtable); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 444 | /* If not, we allocate or mug an existing one: if it's a fresh one, |
| 445 | * repin gets set to 1. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 446 | if (newpgdir == ARRAY_SIZE(lg->pgdirs)) |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 447 | newpgdir = new_pgdir(cpu, pgtable, &repin); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 448 | /* Change the current pgd index to the new one. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 449 | cpu->cpu_pgd = newpgdir; |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 450 | /* If it was completely blank, we map in the Guest kernel stack */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 451 | if (repin) |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 452 | pin_stack_pages(cpu); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 453 | } |
| 454 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 455 | /*H:470 Finally, a routine which throws away everything: all PGD entries in all |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 456 | * the shadow page tables, including the Guest's kernel mappings. This is used |
| 457 | * when we destroy the Guest. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 458 | static void release_all_pagetables(struct lguest *lg) |
| 459 | { |
| 460 | unsigned int i, j; |
| 461 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 462 | /* Every shadow pagetable this Guest has */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 463 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
| 464 | if (lg->pgdirs[i].pgdir) |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 465 | /* Every PGD entry except the Switcher at the top */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 466 | for (j = 0; j < SWITCHER_PGD_INDEX; j++) |
| 467 | release_pgd(lg, lg->pgdirs[i].pgdir + j); |
| 468 | } |
| 469 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 470 | /* We also throw away everything when a Guest tells us it's changed a kernel |
| 471 | * mapping. Since kernel mappings are in every page table, it's easiest to |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 472 | * throw them all away. This traps the Guest in amber for a while as |
| 473 | * everything faults back in, but it's rare. */ |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 474 | void guest_pagetable_clear_all(struct lg_cpu *cpu) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 475 | { |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 476 | release_all_pagetables(cpu->lg); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 477 | /* We need the Guest kernel stack mapped again. */ |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 478 | pin_stack_pages(cpu); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 479 | } |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 480 | /*:*/ |
| 481 | /*M:009 Since we throw away all mappings when a kernel mapping changes, our |
| 482 | * performance sucks for guests using highmem. In fact, a guest with |
| 483 | * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is |
| 484 | * usually slower than a Guest with less memory. |
| 485 | * |
| 486 | * This, of course, cannot be fixed. It would take some kind of... well, I |
| 487 | * don't know, but the term "puissant code-fu" comes to mind. :*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 488 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 489 | /*H:420 This is the routine which actually sets the page table entry for then |
| 490 | * "idx"'th shadow page table. |
| 491 | * |
| 492 | * Normally, we can just throw out the old entry and replace it with 0: if they |
| 493 | * use it demand_page() will put the new entry in. We need to do this anyway: |
| 494 | * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page |
| 495 | * is read from, and _PAGE_DIRTY when it's written to. |
| 496 | * |
| 497 | * But Avi Kivity pointed out that most Operating Systems (Linux included) set |
| 498 | * these bits on PTEs immediately anyway. This is done to save the CPU from |
| 499 | * having to update them, but it helps us the same way: if they set |
| 500 | * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if |
| 501 | * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately. |
| 502 | */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 503 | static void do_set_pte(struct lguest *lg, int idx, |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 504 | unsigned long vaddr, pte_t gpte) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 505 | { |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 506 | /* Look up the matching shadow page directory entry. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 507 | pgd_t *spgd = spgd_addr(lg, idx, vaddr); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 508 | |
| 509 | /* If the top level isn't present, there's no entry to update. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 510 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 511 | /* Otherwise, we start by releasing the existing entry. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 512 | pte_t *spte = spte_addr(lg, *spgd, vaddr); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 513 | release_pte(*spte); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 514 | |
| 515 | /* If they're setting this entry as dirty or accessed, we might |
| 516 | * as well put that entry they've given us in now. This shaves |
| 517 | * 10% off a copy-on-write micro-benchmark. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 518 | if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) { |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 519 | check_gpte(lg, gpte); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 520 | *spte = gpte_to_spte(lg, gpte, |
| 521 | pte_flags(gpte) & _PAGE_DIRTY); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 522 | } else |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 523 | /* Otherwise kill it and we can demand_page() it in |
| 524 | * later. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 525 | *spte = __pte(0); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 526 | } |
| 527 | } |
| 528 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 529 | /*H:410 Updating a PTE entry is a little trickier. |
| 530 | * |
| 531 | * We keep track of several different page tables (the Guest uses one for each |
| 532 | * process, so it makes sense to cache at least a few). Each of these have |
| 533 | * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for |
| 534 | * all processes. So when the page table above that address changes, we update |
| 535 | * all the page tables, not just the current one. This is rare. |
| 536 | * |
| 537 | * The benefit is that when we have to track a new page table, we can copy keep |
| 538 | * all the kernel mappings. This speeds up context switch immensely. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 539 | void guest_set_pte(struct lguest *lg, |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 540 | unsigned long gpgdir, unsigned long vaddr, pte_t gpte) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 541 | { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 542 | /* Kernel mappings must be changed on all top levels. Slow, but |
| 543 | * doesn't happen often. */ |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 544 | if (vaddr >= lg->kernel_address) { |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 545 | unsigned int i; |
| 546 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
| 547 | if (lg->pgdirs[i].pgdir) |
| 548 | do_set_pte(lg, i, vaddr, gpte); |
| 549 | } else { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 550 | /* Is this page table one we have a shadow for? */ |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 551 | int pgdir = find_pgdir(lg, gpgdir); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 552 | if (pgdir != ARRAY_SIZE(lg->pgdirs)) |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 553 | /* If so, do the update. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 554 | do_set_pte(lg, pgdir, vaddr, gpte); |
| 555 | } |
| 556 | } |
| 557 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 558 | /*H:400 |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 559 | * (iii) Setting up a page table entry when the Guest tells us one has changed. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 560 | * |
| 561 | * Just like we did in interrupts_and_traps.c, it makes sense for us to deal |
| 562 | * with the other side of page tables while we're here: what happens when the |
| 563 | * Guest asks for a page table to be updated? |
| 564 | * |
| 565 | * We already saw that demand_page() will fill in the shadow page tables when |
| 566 | * needed, so we can simply remove shadow page table entries whenever the Guest |
| 567 | * tells us they've changed. When the Guest tries to use the new entry it will |
| 568 | * fault and demand_page() will fix it up. |
| 569 | * |
| 570 | * So with that in mind here's our code to to update a (top-level) PGD entry: |
| 571 | */ |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 572 | void guest_set_pmd(struct lguest *lg, unsigned long gpgdir, u32 idx) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 573 | { |
| 574 | int pgdir; |
| 575 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 576 | /* The kernel seems to try to initialize this early on: we ignore its |
| 577 | * attempts to map over the Switcher. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 578 | if (idx >= SWITCHER_PGD_INDEX) |
| 579 | return; |
| 580 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 581 | /* If they're talking about a page table we have a shadow for... */ |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 582 | pgdir = find_pgdir(lg, gpgdir); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 583 | if (pgdir < ARRAY_SIZE(lg->pgdirs)) |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 584 | /* ... throw it away. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 585 | release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx); |
| 586 | } |
| 587 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 588 | /*H:500 (vii) Setting up the page tables initially. |
| 589 | * |
| 590 | * When a Guest is first created, the Launcher tells us where the toplevel of |
| 591 | * its first page table is. We set some things up here: */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 592 | int init_guest_pagetable(struct lguest *lg, unsigned long pgtable) |
| 593 | { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 594 | /* We start on the first shadow page table, and give it a blank PGD |
| 595 | * page. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 596 | lg->pgdirs[0].gpgdir = pgtable; |
| 597 | lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL); |
| 598 | if (!lg->pgdirs[0].pgdir) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 599 | return -ENOMEM; |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 600 | lg->cpus[0].cpu_pgd = 0; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 601 | return 0; |
| 602 | } |
| 603 | |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 604 | /* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ |
| 605 | void page_table_guest_data_init(struct lguest *lg) |
| 606 | { |
| 607 | /* We get the kernel address: above this is all kernel memory. */ |
| 608 | if (get_user(lg->kernel_address, &lg->lguest_data->kernel_address) |
| 609 | /* We tell the Guest that it can't use the top 4MB of virtual |
| 610 | * addresses used by the Switcher. */ |
| 611 | || put_user(4U*1024*1024, &lg->lguest_data->reserve_mem) |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 612 | || put_user(lg->pgdirs[0].gpgdir, &lg->lguest_data->pgdir)) |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 613 | kill_guest(lg, "bad guest page %p", lg->lguest_data); |
| 614 | |
| 615 | /* In flush_user_mappings() we loop from 0 to |
| 616 | * "pgd_index(lg->kernel_address)". This assumes it won't hit the |
| 617 | * Switcher mappings, so check that now. */ |
| 618 | if (pgd_index(lg->kernel_address) >= SWITCHER_PGD_INDEX) |
| 619 | kill_guest(lg, "bad kernel address %#lx", lg->kernel_address); |
| 620 | } |
| 621 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 622 | /* When a Guest dies, our cleanup is fairly simple. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 623 | void free_guest_pagetable(struct lguest *lg) |
| 624 | { |
| 625 | unsigned int i; |
| 626 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 627 | /* Throw away all page table pages. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 628 | release_all_pagetables(lg); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 629 | /* Now free the top levels: free_page() can handle 0 just fine. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 630 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
| 631 | free_page((long)lg->pgdirs[i].pgdir); |
| 632 | } |
| 633 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 634 | /*H:480 (vi) Mapping the Switcher when the Guest is about to run. |
| 635 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 636 | * The Switcher and the two pages for this CPU need to be visible in the |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 637 | * Guest (and not the pages for other CPUs). We have the appropriate PTE pages |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 638 | * for each CPU already set up, we just need to hook them in now we know which |
| 639 | * Guest is about to run on this CPU. */ |
Glauber de Oliveira Costa | 0c78441 | 2008-01-07 11:05:30 -0200 | [diff] [blame] | 640 | void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 641 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 642 | pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages); |
| 643 | pgd_t switcher_pgd; |
| 644 | pte_t regs_pte; |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 645 | unsigned long pfn; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 646 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 647 | /* Make the last PGD entry for this Guest point to the Switcher's PTE |
| 648 | * page for this CPU (with appropriate flags). */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 649 | switcher_pgd = __pgd(__pa(switcher_pte_page) | _PAGE_KERNEL); |
| 650 | |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame^] | 651 | cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 652 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 653 | /* We also change the Switcher PTE page. When we're running the Guest, |
| 654 | * we want the Guest's "regs" page to appear where the first Switcher |
| 655 | * page for this CPU is. This is an optimization: when the Switcher |
| 656 | * saves the Guest registers, it saves them into the first page of this |
| 657 | * CPU's "struct lguest_pages": if we make sure the Guest's register |
| 658 | * page is already mapped there, we don't have to copy them out |
| 659 | * again. */ |
Glauber de Oliveira Costa | a53a35a | 2008-01-07 11:05:32 -0200 | [diff] [blame] | 660 | pfn = __pa(cpu->regs_page) >> PAGE_SHIFT; |
| 661 | regs_pte = pfn_pte(pfn, __pgprot(_PAGE_KERNEL)); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 662 | switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTRS_PER_PTE] = regs_pte; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 663 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 664 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 665 | |
| 666 | static void free_switcher_pte_pages(void) |
| 667 | { |
| 668 | unsigned int i; |
| 669 | |
| 670 | for_each_possible_cpu(i) |
| 671 | free_page((long)switcher_pte_page(i)); |
| 672 | } |
| 673 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 674 | /*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given |
| 675 | * the CPU number and the "struct page"s for the Switcher code itself. |
| 676 | * |
| 677 | * Currently the Switcher is less than a page long, so "pages" is always 1. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 678 | static __init void populate_switcher_pte_page(unsigned int cpu, |
| 679 | struct page *switcher_page[], |
| 680 | unsigned int pages) |
| 681 | { |
| 682 | unsigned int i; |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 683 | pte_t *pte = switcher_pte_page(cpu); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 684 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 685 | /* The first entries are easy: they map the Switcher code. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 686 | for (i = 0; i < pages; i++) { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 687 | pte[i] = mk_pte(switcher_page[i], |
| 688 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 689 | } |
| 690 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 691 | /* The only other thing we map is this CPU's pair of pages. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 692 | i = pages + cpu*2; |
| 693 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 694 | /* First page (Guest registers) is writable from the Guest */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 695 | pte[i] = pfn_pte(page_to_pfn(switcher_page[i]), |
| 696 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW)); |
| 697 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 698 | /* The second page contains the "struct lguest_ro_state", and is |
| 699 | * read-only. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 700 | pte[i+1] = pfn_pte(page_to_pfn(switcher_page[i+1]), |
| 701 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 702 | } |
| 703 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 704 | /* We've made it through the page table code. Perhaps our tired brains are |
| 705 | * still processing the details, or perhaps we're simply glad it's over. |
| 706 | * |
| 707 | * If nothing else, note that all this complexity in juggling shadow page |
| 708 | * tables in sync with the Guest's page tables is for one reason: for most |
| 709 | * Guests this page table dance determines how bad performance will be. This |
| 710 | * is why Xen uses exotic direct Guest pagetable manipulation, and why both |
| 711 | * Intel and AMD have implemented shadow page table support directly into |
| 712 | * hardware. |
| 713 | * |
| 714 | * There is just one file remaining in the Host. */ |
| 715 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 716 | /*H:510 At boot or module load time, init_pagetables() allocates and populates |
| 717 | * the Switcher PTE page for each CPU. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 718 | __init int init_pagetables(struct page **switcher_page, unsigned int pages) |
| 719 | { |
| 720 | unsigned int i; |
| 721 | |
| 722 | for_each_possible_cpu(i) { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 723 | switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 724 | if (!switcher_pte_page(i)) { |
| 725 | free_switcher_pte_pages(); |
| 726 | return -ENOMEM; |
| 727 | } |
| 728 | populate_switcher_pte_page(i, switcher_page, pages); |
| 729 | } |
| 730 | return 0; |
| 731 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 732 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 733 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 734 | /* Cleaning up simply involves freeing the PTE page for each CPU. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 735 | void free_pagetables(void) |
| 736 | { |
| 737 | free_switcher_pte_pages(); |
| 738 | } |