Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1 | /*P:700 |
| 2 | * The pagetable code, on the other hand, still shows the scars of |
Rusty Russell | f938d2c | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 3 | * previous encounters. It's functional, and as neat as it can be in the |
| 4 | * circumstances, but be wary, for these things are subtle and break easily. |
| 5 | * The Guest provides a virtual to physical mapping, but we can neither trust |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 6 | * it nor use it: we verify and convert it here then point the CPU to the |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 7 | * converted Guest pages when running the Guest. |
| 8 | :*/ |
Rusty Russell | f938d2c | 2007-07-26 10:41:02 -0700 | [diff] [blame] | 9 | |
| 10 | /* Copyright (C) Rusty Russell IBM Corporation 2006. |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 11 | * GPL v2 and any later version */ |
| 12 | #include <linux/mm.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 13 | #include <linux/gfp.h> |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 14 | #include <linux/types.h> |
| 15 | #include <linux/spinlock.h> |
| 16 | #include <linux/random.h> |
| 17 | #include <linux/percpu.h> |
| 18 | #include <asm/tlbflush.h> |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 19 | #include <asm/uaccess.h> |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 20 | #include "lg.h" |
| 21 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 22 | /*M:008 |
| 23 | * We hold reference to pages, which prevents them from being swapped. |
Rusty Russell | f56a384 | 2007-07-26 10:41:05 -0700 | [diff] [blame] | 24 | * It'd be nice to have a callback in the "struct mm_struct" when Linux wants |
| 25 | * to swap out. If we had this, and a shrinker callback to trim PTE pages, we |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 26 | * could probably consider launching Guests as non-root. |
| 27 | :*/ |
Rusty Russell | f56a384 | 2007-07-26 10:41:05 -0700 | [diff] [blame] | 28 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 29 | /*H:300 |
| 30 | * The Page Table Code |
| 31 | * |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 32 | * We use two-level page tables for the Guest, or three-level with PAE. If |
| 33 | * you're not entirely comfortable with virtual addresses, physical addresses |
| 34 | * and page tables then I recommend you review arch/x86/lguest/boot.c's "Page |
| 35 | * Table Handling" (with diagrams!). |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 36 | * |
| 37 | * The Guest keeps page tables, but we maintain the actual ones here: these are |
| 38 | * called "shadow" page tables. Which is a very Guest-centric name: these are |
| 39 | * the real page tables the CPU uses, although we keep them up to date to |
| 40 | * reflect the Guest's. (See what I mean about weird naming? Since when do |
| 41 | * shadows reflect anything?) |
| 42 | * |
| 43 | * Anyway, this is the most complicated part of the Host code. There are seven |
| 44 | * parts to this: |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 45 | * (i) Looking up a page table entry when the Guest faults, |
| 46 | * (ii) Making sure the Guest stack is mapped, |
| 47 | * (iii) Setting up a page table entry when the Guest tells us one has changed, |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 48 | * (iv) Switching page tables, |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 49 | * (v) Flushing (throwing away) page tables, |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 50 | * (vi) Mapping the Switcher when the Guest is about to run, |
| 51 | * (vii) Setting up the page tables initially. |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 52 | :*/ |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 53 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 54 | /* |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 55 | * The Switcher uses the complete top PTE page. That's 1024 PTE entries (4MB) |
| 56 | * or 512 PTE entries with PAE (2MB). |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 57 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 58 | #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 59 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 60 | /* |
| 61 | * For PAE we need the PMD index as well. We use the last 2MB, so we |
| 62 | * will need the last pmd entry of the last pmd page. |
| 63 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 64 | #ifdef CONFIG_X86_PAE |
| 65 | #define SWITCHER_PMD_INDEX (PTRS_PER_PMD - 1) |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 66 | #define CHECK_GPGD_MASK _PAGE_PRESENT |
| 67 | #else |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 68 | #define CHECK_GPGD_MASK _PAGE_TABLE |
| 69 | #endif |
| 70 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 71 | /* |
| 72 | * We actually need a separate PTE page for each CPU. Remember that after the |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 73 | * Switcher code itself comes two pages for each CPU, and we don't want this |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 74 | * CPU's guest to see the pages of any other CPU. |
| 75 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 76 | static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 77 | #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) |
| 78 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 79 | /*H:320 |
| 80 | * The page table code is curly enough to need helper functions to keep it |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 81 | * clear and clean. The kernel itself provides many of them; one advantage |
| 82 | * of insisting that the Guest and Host use the same CONFIG_PAE setting. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 83 | * |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 84 | * There are two functions which return pointers to the shadow (aka "real") |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 85 | * page tables. |
| 86 | * |
| 87 | * 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] | 88 | * page directory entry (PGD) for that address. Since we keep track of several |
| 89 | * page tables, the "i" argument tells us which one we're interested in (it's |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 90 | * usually the current one). |
| 91 | */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 92 | static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 93 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 94 | unsigned int index = pgd_index(vaddr); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 95 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 96 | /* Return a pointer index'th pgd entry for the i'th page table. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 97 | return &cpu->lg->pgdirs[i].pgdir[index]; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 98 | } |
| 99 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 100 | #ifdef CONFIG_X86_PAE |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 101 | /* |
| 102 | * This routine then takes the PGD entry given above, which contains the |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 103 | * address of the PMD page. It then returns a pointer to the PMD entry for the |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 104 | * given address. |
| 105 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 106 | static pmd_t *spmd_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) |
| 107 | { |
| 108 | unsigned int index = pmd_index(vaddr); |
| 109 | pmd_t *page; |
| 110 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 111 | /* You should never call this if the PGD entry wasn't valid */ |
| 112 | BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); |
| 113 | page = __va(pgd_pfn(spgd) << PAGE_SHIFT); |
| 114 | |
| 115 | return &page[index]; |
| 116 | } |
| 117 | #endif |
| 118 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 119 | /* |
| 120 | * This routine then takes the page directory entry returned above, which |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 121 | * contains the address of the page table entry (PTE) page. It then returns a |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 122 | * pointer to the PTE entry for the given address. |
| 123 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 124 | static pte_t *spte_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 125 | { |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 126 | #ifdef CONFIG_X86_PAE |
| 127 | pmd_t *pmd = spmd_addr(cpu, spgd, vaddr); |
| 128 | pte_t *page = __va(pmd_pfn(*pmd) << PAGE_SHIFT); |
| 129 | |
| 130 | /* You should never call this if the PMD entry wasn't valid */ |
| 131 | BUG_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)); |
| 132 | #else |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 133 | pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 134 | /* You should never call this if the PGD entry wasn't valid */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 135 | BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 136 | #endif |
| 137 | |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 138 | return &page[pte_index(vaddr)]; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 139 | } |
| 140 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 141 | /* |
Rusty Russell | 9f54288 | 2011-07-22 14:39:50 +0930 | [diff] [blame] | 142 | * These functions are just like the above, except they access the Guest |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 143 | * page tables. Hence they return a Guest address. |
| 144 | */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 145 | static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 146 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 147 | unsigned int index = vaddr >> (PGDIR_SHIFT); |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 148 | return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 149 | } |
| 150 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 151 | #ifdef CONFIG_X86_PAE |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 152 | /* Follow the PGD to the PMD. */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 153 | static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 154 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 155 | unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; |
| 156 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 157 | return gpage + pmd_index(vaddr) * sizeof(pmd_t); |
| 158 | } |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 159 | |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 160 | /* Follow the PMD to the PTE. */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 161 | static unsigned long gpte_addr(struct lg_cpu *cpu, |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 162 | pmd_t gpmd, unsigned long vaddr) |
| 163 | { |
| 164 | unsigned long gpage = pmd_pfn(gpmd) << PAGE_SHIFT; |
| 165 | |
| 166 | BUG_ON(!(pmd_flags(gpmd) & _PAGE_PRESENT)); |
| 167 | return gpage + pte_index(vaddr) * sizeof(pte_t); |
| 168 | } |
| 169 | #else |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 170 | /* Follow the PGD to the PTE (no mid-level for !PAE). */ |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 171 | static unsigned long gpte_addr(struct lg_cpu *cpu, |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 172 | pgd_t gpgd, unsigned long vaddr) |
| 173 | { |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 174 | unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 175 | |
| 176 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 177 | return gpage + pte_index(vaddr) * sizeof(pte_t); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 178 | } |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 179 | #endif |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 180 | /*:*/ |
| 181 | |
Rusty Russell | 9f54288 | 2011-07-22 14:39:50 +0930 | [diff] [blame] | 182 | /*M:007 |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 183 | * get_pfn is slow: we could probably try to grab batches of pages here as |
| 184 | * an optimization (ie. pre-faulting). |
| 185 | :*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 186 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 187 | /*H:350 |
| 188 | * This routine takes a page number given by the Guest and converts it to |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 189 | * an actual, physical page number. It can fail for several reasons: the |
| 190 | * virtual address might not be mapped by the Launcher, the write flag is set |
| 191 | * and the page is read-only, or the write flag was set and the page was |
| 192 | * shared so had to be copied, but we ran out of memory. |
| 193 | * |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 194 | * This holds a reference to the page, so release_pte() is careful to put that |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 195 | * back. |
| 196 | */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 197 | static unsigned long get_pfn(unsigned long virtpfn, int write) |
| 198 | { |
| 199 | struct page *page; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 200 | |
Rusty Russell | 71a3f4e | 2008-08-12 17:52:53 -0500 | [diff] [blame] | 201 | /* gup me one page at this address please! */ |
| 202 | if (get_user_pages_fast(virtpfn << PAGE_SHIFT, 1, write, &page) == 1) |
| 203 | return page_to_pfn(page); |
| 204 | |
| 205 | /* This value indicates failure. */ |
| 206 | return -1UL; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 207 | } |
| 208 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 209 | /*H:340 |
| 210 | * Converting a Guest page table entry to a shadow (ie. real) page table |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 211 | * entry can be a little tricky. The flags are (almost) the same, but the |
| 212 | * Guest PTE contains a virtual page number: the CPU needs the real page |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 213 | * number. |
| 214 | */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 215 | static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 216 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 217 | unsigned long pfn, base, flags; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 218 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 219 | /* |
| 220 | * The Guest sets the global flag, because it thinks that it is using |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 221 | * PGE. We only told it to use PGE so it would tell us whether it was |
| 222 | * flushing a kernel mapping or a userspace mapping. We don't actually |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 223 | * use the global bit, so throw it away. |
| 224 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 225 | flags = (pte_flags(gpte) & ~_PAGE_GLOBAL); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 226 | |
Rusty Russell | 3c6b5bf | 2007-10-22 11:03:26 +1000 | [diff] [blame] | 227 | /* The Guest's pages are offset inside the Launcher. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 228 | base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE; |
Rusty Russell | 3c6b5bf | 2007-10-22 11:03:26 +1000 | [diff] [blame] | 229 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 230 | /* |
| 231 | * We need a temporary "unsigned long" variable to hold the answer from |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 232 | * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't |
| 233 | * fit in spte.pfn. get_pfn() finds the real physical number of the |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 234 | * page, given the virtual number. |
| 235 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 236 | pfn = get_pfn(base + pte_pfn(gpte), write); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 237 | if (pfn == -1UL) { |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 238 | kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte)); |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 239 | /* |
| 240 | * When we destroy the Guest, we'll go through the shadow page |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 241 | * tables and release_pte() them. Make sure we don't think |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 242 | * this one is valid! |
| 243 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 244 | flags = 0; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 245 | } |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 246 | /* Now we assemble our shadow PTE from the page number and flags. */ |
| 247 | return pfn_pte(pfn, __pgprot(flags)); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 248 | } |
| 249 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 250 | /*H:460 And to complete the chain, release_pte() looks like this: */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 251 | static void release_pte(pte_t pte) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 252 | { |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 253 | /* |
| 254 | * Remember that get_user_pages_fast() took a reference to the page, in |
| 255 | * get_pfn()? We have to put it back now. |
| 256 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 257 | if (pte_flags(pte) & _PAGE_PRESENT) |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 258 | put_page(pte_page(pte)); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 259 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 260 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 261 | |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 262 | static bool check_gpte(struct lg_cpu *cpu, pte_t gpte) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 263 | { |
Ahmed S. Darwish | 31f4b46 | 2008-02-09 23:24:09 +0100 | [diff] [blame] | 264 | if ((pte_flags(gpte) & _PAGE_PSE) || |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 265 | pte_pfn(gpte) >= cpu->lg->pfn_limit) { |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 266 | kill_guest(cpu, "bad page table entry"); |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 267 | return false; |
| 268 | } |
| 269 | return true; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 270 | } |
| 271 | |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 272 | static bool check_gpgd(struct lg_cpu *cpu, pgd_t gpgd) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 273 | { |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 274 | if ((pgd_flags(gpgd) & ~CHECK_GPGD_MASK) || |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 275 | (pgd_pfn(gpgd) >= cpu->lg->pfn_limit)) { |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 276 | kill_guest(cpu, "bad page directory entry"); |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 277 | return false; |
| 278 | } |
| 279 | return true; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 280 | } |
| 281 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 282 | #ifdef CONFIG_X86_PAE |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 283 | static bool check_gpmd(struct lg_cpu *cpu, pmd_t gpmd) |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 284 | { |
| 285 | if ((pmd_flags(gpmd) & ~_PAGE_TABLE) || |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 286 | (pmd_pfn(gpmd) >= cpu->lg->pfn_limit)) { |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 287 | kill_guest(cpu, "bad page middle directory entry"); |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 288 | return false; |
| 289 | } |
| 290 | return true; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 291 | } |
| 292 | #endif |
| 293 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 294 | /*H:330 |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 295 | * (i) Looking up a page table entry when the Guest faults. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 296 | * |
| 297 | * We saw this call in run_guest(): when we see a page fault in the Guest, we |
| 298 | * come here. That's because we only set up the shadow page tables lazily as |
| 299 | * they're needed, so we get page faults all the time and quietly fix them up |
| 300 | * and return to the Guest without it knowing. |
| 301 | * |
| 302 | * If we fixed up the fault (ie. we mapped the address), this routine returns |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 303 | * true. Otherwise, it was a real fault and we need to tell the Guest. |
| 304 | */ |
Matias Zabaljauregui | df1693a | 2009-03-18 13:38:35 -0300 | [diff] [blame] | 305 | bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 306 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 307 | pgd_t gpgd; |
| 308 | pgd_t *spgd; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 309 | unsigned long gpte_ptr; |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 310 | pte_t gpte; |
| 311 | pte_t *spte; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 312 | |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 313 | /* Mid level for PAE. */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 314 | #ifdef CONFIG_X86_PAE |
| 315 | pmd_t *spmd; |
| 316 | pmd_t gpmd; |
| 317 | #endif |
| 318 | |
Rusty Russell | 68a644d | 2013-04-22 14:10:37 +0930 | [diff] [blame] | 319 | /* We never demand page the Switcher, so trying is a mistake. */ |
| 320 | if (vaddr >= switcher_addr) |
| 321 | return false; |
| 322 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 323 | /* First step: get the top-level Guest page table entry. */ |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 324 | if (unlikely(cpu->linear_pages)) { |
| 325 | /* Faking up a linear mapping. */ |
| 326 | gpgd = __pgd(CHECK_GPGD_MASK); |
| 327 | } else { |
| 328 | gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); |
| 329 | /* Toplevel not present? We can't map it in. */ |
| 330 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) |
| 331 | return false; |
| 332 | } |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 333 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 334 | /* Now look at the matching shadow entry. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 335 | spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 336 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 337 | /* No shadow entry: allocate a new shadow PTE page. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 338 | unsigned long ptepage = get_zeroed_page(GFP_KERNEL); |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 339 | /* |
| 340 | * This is not really the Guest's fault, but killing it is |
| 341 | * simple for this corner case. |
| 342 | */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 343 | if (!ptepage) { |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 344 | kill_guest(cpu, "out of memory allocating pte page"); |
Matias Zabaljauregui | df1693a | 2009-03-18 13:38:35 -0300 | [diff] [blame] | 345 | return false; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 346 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 347 | /* We check that the Guest pgd is OK. */ |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 348 | if (!check_gpgd(cpu, gpgd)) |
| 349 | return false; |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 350 | /* |
| 351 | * And we copy the flags to the shadow PGD entry. The page |
| 352 | * number in the shadow PGD is the page we just allocated. |
| 353 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 354 | set_pgd(spgd, __pgd(__pa(ptepage) | pgd_flags(gpgd))); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 355 | } |
| 356 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 357 | #ifdef CONFIG_X86_PAE |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 358 | if (unlikely(cpu->linear_pages)) { |
| 359 | /* Faking up a linear mapping. */ |
| 360 | gpmd = __pmd(_PAGE_TABLE); |
| 361 | } else { |
| 362 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); |
| 363 | /* Middle level not present? We can't map it in. */ |
| 364 | if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) |
| 365 | return false; |
| 366 | } |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 367 | |
| 368 | /* Now look at the matching shadow entry. */ |
| 369 | spmd = spmd_addr(cpu, *spgd, vaddr); |
| 370 | |
| 371 | if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) { |
| 372 | /* No shadow entry: allocate a new shadow PTE page. */ |
| 373 | unsigned long ptepage = get_zeroed_page(GFP_KERNEL); |
| 374 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 375 | /* |
| 376 | * This is not really the Guest's fault, but killing it is |
| 377 | * simple for this corner case. |
| 378 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 379 | if (!ptepage) { |
| 380 | kill_guest(cpu, "out of memory allocating pte page"); |
| 381 | return false; |
| 382 | } |
| 383 | |
| 384 | /* We check that the Guest pmd is OK. */ |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 385 | if (!check_gpmd(cpu, gpmd)) |
| 386 | return false; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 387 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 388 | /* |
| 389 | * And we copy the flags to the shadow PMD entry. The page |
| 390 | * number in the shadow PMD is the page we just allocated. |
| 391 | */ |
Rusty Russell | 4c1ea3d | 2009-09-23 22:26:45 -0600 | [diff] [blame] | 392 | set_pmd(spmd, __pmd(__pa(ptepage) | pmd_flags(gpmd))); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 393 | } |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 394 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 395 | /* |
| 396 | * OK, now we look at the lower level in the Guest page table: keep its |
| 397 | * address, because we might update it later. |
| 398 | */ |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 399 | gpte_ptr = gpte_addr(cpu, gpmd, vaddr); |
| 400 | #else |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 401 | /* |
| 402 | * OK, now we look at the lower level in the Guest page table: keep its |
| 403 | * address, because we might update it later. |
| 404 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 405 | gpte_ptr = gpte_addr(cpu, gpgd, vaddr); |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 406 | #endif |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 407 | |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 408 | if (unlikely(cpu->linear_pages)) { |
| 409 | /* Linear? Make up a PTE which points to same page. */ |
| 410 | gpte = __pte((vaddr & PAGE_MASK) | _PAGE_RW | _PAGE_PRESENT); |
| 411 | } else { |
| 412 | /* Read the actual PTE value. */ |
| 413 | gpte = lgread(cpu, gpte_ptr, pte_t); |
| 414 | } |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 415 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 416 | /* 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] | 417 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
Matias Zabaljauregui | df1693a | 2009-03-18 13:38:35 -0300 | [diff] [blame] | 418 | return false; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 419 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 420 | /* |
| 421 | * Check they're not trying to write to a page the Guest wants |
| 422 | * read-only (bit 2 of errcode == write). |
| 423 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 424 | if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW)) |
Matias Zabaljauregui | df1693a | 2009-03-18 13:38:35 -0300 | [diff] [blame] | 425 | return false; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 426 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 427 | /* User access to a kernel-only page? (bit 3 == user access) */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 428 | if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER)) |
Matias Zabaljauregui | df1693a | 2009-03-18 13:38:35 -0300 | [diff] [blame] | 429 | return false; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 430 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 431 | /* |
| 432 | * Check that the Guest PTE flags are OK, and the page number is below |
| 433 | * the pfn_limit (ie. not mapping the Launcher binary). |
| 434 | */ |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 435 | if (!check_gpte(cpu, gpte)) |
| 436 | return false; |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 437 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 438 | /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 439 | gpte = pte_mkyoung(gpte); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 440 | if (errcode & 2) |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 441 | gpte = pte_mkdirty(gpte); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 442 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 443 | /* Get the pointer to the shadow PTE entry we're going to set. */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 444 | spte = spte_addr(cpu, *spgd, vaddr); |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 445 | |
| 446 | /* |
| 447 | * If there was a valid shadow PTE entry here before, we release it. |
| 448 | * This can happen with a write to a previously read-only entry. |
| 449 | */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 450 | release_pte(*spte); |
| 451 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 452 | /* |
| 453 | * If this is a write, we insist that the Guest page is writable (the |
| 454 | * final arg to gpte_to_spte()). |
| 455 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 456 | if (pte_dirty(gpte)) |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 457 | *spte = gpte_to_spte(cpu, gpte, 1); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 458 | else |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 459 | /* |
| 460 | * If this is a read, don't set the "writable" bit in the page |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 461 | * table entry, even if the Guest says it's writable. That way |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 462 | * we will come back here when a write does actually occur, so |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 463 | * we can update the Guest's _PAGE_DIRTY flag. |
| 464 | */ |
Rusty Russell | 4c1ea3d | 2009-09-23 22:26:45 -0600 | [diff] [blame] | 465 | set_pte(spte, gpte_to_spte(cpu, pte_wrprotect(gpte), 0)); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 466 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 467 | /* |
| 468 | * Finally, we write the Guest PTE entry back: we've set the |
| 469 | * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. |
| 470 | */ |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 471 | if (likely(!cpu->linear_pages)) |
| 472 | lgwrite(cpu, gpte_ptr, pte_t, gpte); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 473 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 474 | /* |
| 475 | * The fault is fixed, the page table is populated, the mapping |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 476 | * manipulated, the result returned and the code complete. A small |
| 477 | * delay and a trace of alliteration are the only indications the Guest |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 478 | * has that a page fault occurred at all. |
| 479 | */ |
Matias Zabaljauregui | df1693a | 2009-03-18 13:38:35 -0300 | [diff] [blame] | 480 | return true; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 481 | } |
| 482 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 483 | /*H:360 |
| 484 | * (ii) Making sure the Guest stack is mapped. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 485 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 486 | * Remember that direct traps into the Guest need a mapped Guest kernel stack. |
| 487 | * pin_stack_pages() calls us here: we could simply call demand_page(), but as |
| 488 | * we've seen that logic is quite long, and usually the stack pages are already |
| 489 | * mapped, so it's overkill. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 490 | * |
| 491 | * This is a quick version which answers the question: is this virtual address |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 492 | * mapped by the shadow page tables, and is it writable? |
| 493 | */ |
Matias Zabaljauregui | df1693a | 2009-03-18 13:38:35 -0300 | [diff] [blame] | 494 | static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 495 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 496 | pgd_t *spgd; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 497 | unsigned long flags; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 498 | #ifdef CONFIG_X86_PAE |
| 499 | pmd_t *spmd; |
| 500 | #endif |
Rusty Russell | 68a644d | 2013-04-22 14:10:37 +0930 | [diff] [blame] | 501 | |
| 502 | /* You can't put your stack in the Switcher! */ |
| 503 | if (vaddr >= switcher_addr) |
| 504 | return false; |
| 505 | |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 506 | /* Look at the current top level entry: is it present? */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 507 | spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 508 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) |
Matias Zabaljauregui | df1693a | 2009-03-18 13:38:35 -0300 | [diff] [blame] | 509 | return false; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 510 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 511 | #ifdef CONFIG_X86_PAE |
| 512 | spmd = spmd_addr(cpu, *spgd, vaddr); |
| 513 | if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) |
| 514 | return false; |
| 515 | #endif |
| 516 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 517 | /* |
| 518 | * Check the flags on the pte entry itself: it must be present and |
| 519 | * writable. |
| 520 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 521 | flags = pte_flags(*(spte_addr(cpu, *spgd, vaddr))); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 522 | |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 523 | return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); |
| 524 | } |
| 525 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 526 | /* |
| 527 | * So, when pin_stack_pages() asks us to pin a page, we check if it's already |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 528 | * in the page tables, and if not, we call demand_page() with error code 2 |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 529 | * (meaning "write"). |
| 530 | */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 531 | void pin_page(struct lg_cpu *cpu, unsigned long vaddr) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 532 | { |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 533 | if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2)) |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 534 | kill_guest(cpu, "bad stack page %#lx", vaddr); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 535 | } |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 536 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 537 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 538 | #ifdef CONFIG_X86_PAE |
| 539 | static void release_pmd(pmd_t *spmd) |
| 540 | { |
| 541 | /* If the entry's not present, there's nothing to release. */ |
| 542 | if (pmd_flags(*spmd) & _PAGE_PRESENT) { |
| 543 | unsigned int i; |
| 544 | pte_t *ptepage = __va(pmd_pfn(*spmd) << PAGE_SHIFT); |
| 545 | /* For each entry in the page, we might need to release it. */ |
| 546 | for (i = 0; i < PTRS_PER_PTE; i++) |
| 547 | release_pte(ptepage[i]); |
| 548 | /* Now we can free the page of PTEs */ |
| 549 | free_page((long)ptepage); |
| 550 | /* And zero out the PMD entry so we never release it twice. */ |
Rusty Russell | 4c1ea3d | 2009-09-23 22:26:45 -0600 | [diff] [blame] | 551 | set_pmd(spmd, __pmd(0)); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 552 | } |
| 553 | } |
| 554 | |
| 555 | static void release_pgd(pgd_t *spgd) |
| 556 | { |
| 557 | /* If the entry's not present, there's nothing to release. */ |
| 558 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
| 559 | unsigned int i; |
| 560 | pmd_t *pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); |
| 561 | |
| 562 | for (i = 0; i < PTRS_PER_PMD; i++) |
| 563 | release_pmd(&pmdpage[i]); |
| 564 | |
| 565 | /* Now we can free the page of PMDs */ |
| 566 | free_page((long)pmdpage); |
| 567 | /* And zero out the PGD entry so we never release it twice. */ |
| 568 | set_pgd(spgd, __pgd(0)); |
| 569 | } |
| 570 | } |
| 571 | |
| 572 | #else /* !CONFIG_X86_PAE */ |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 573 | /*H:450 |
| 574 | * If we chase down the release_pgd() code, the non-PAE version looks like |
| 575 | * this. The PAE version is almost identical, but instead of calling |
| 576 | * release_pte it calls release_pmd(), which looks much like this. |
| 577 | */ |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 578 | static void release_pgd(pgd_t *spgd) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 579 | { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 580 | /* If the entry's not present, there's nothing to release. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 581 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 582 | unsigned int i; |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 583 | /* |
| 584 | * Converting the pfn to find the actual PTE page is easy: turn |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 585 | * the page number into a physical address, then convert to a |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 586 | * virtual address (easy for kernel pages like this one). |
| 587 | */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 588 | pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 589 | /* For each entry in the page, we might need to release it. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 590 | for (i = 0; i < PTRS_PER_PTE; i++) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 591 | release_pte(ptepage[i]); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 592 | /* Now we can free the page of PTEs */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 593 | free_page((long)ptepage); |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 594 | /* And zero out the PGD entry so we never release it twice. */ |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 595 | *spgd = __pgd(0); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 596 | } |
| 597 | } |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 598 | #endif |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 599 | |
| 600 | /*H:445 |
| 601 | * We saw flush_user_mappings() twice: once from the flush_user_mappings() |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 602 | * hypercall and once in new_pgdir() when we re-used a top-level pgdir page. |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 603 | * It simply releases every PTE page from 0 up to the Guest's kernel address. |
| 604 | */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 605 | static void flush_user_mappings(struct lguest *lg, int idx) |
| 606 | { |
| 607 | unsigned int i; |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 608 | /* Release every pgd entry up to the kernel's address. */ |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 609 | for (i = 0; i < pgd_index(lg->kernel_address); i++) |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 610 | release_pgd(lg->pgdirs[idx].pgdir + i); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 611 | } |
| 612 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 613 | /*H:440 |
| 614 | * (v) Flushing (throwing away) page tables, |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 615 | * |
| 616 | * The Guest has a hypercall to throw away the page tables: it's used when a |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 617 | * large number of mappings have been changed. |
| 618 | */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 619 | void guest_pagetable_flush_user(struct lg_cpu *cpu) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 620 | { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 621 | /* Drop the userspace part of the current page table. */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 622 | flush_user_mappings(cpu->lg, cpu->cpu_pgd); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 623 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 624 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 625 | |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 626 | /* 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] | 627 | unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr) |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 628 | { |
| 629 | pgd_t gpgd; |
| 630 | pte_t gpte; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 631 | #ifdef CONFIG_X86_PAE |
| 632 | pmd_t gpmd; |
| 633 | #endif |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 634 | |
| 635 | /* Still not set up? Just map 1:1. */ |
| 636 | if (unlikely(cpu->linear_pages)) |
| 637 | return vaddr; |
| 638 | |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 639 | /* First step: get the top-level Guest page table entry. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 640 | gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 641 | /* Toplevel not present? We can't map it in. */ |
Rusty Russell | 6afbdd0 | 2009-03-30 21:55:23 -0600 | [diff] [blame] | 642 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) { |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 643 | kill_guest(cpu, "Bad address %#lx", vaddr); |
Rusty Russell | 6afbdd0 | 2009-03-30 21:55:23 -0600 | [diff] [blame] | 644 | return -1UL; |
| 645 | } |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 646 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 647 | #ifdef CONFIG_X86_PAE |
| 648 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); |
| 649 | if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) |
| 650 | kill_guest(cpu, "Bad address %#lx", vaddr); |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 651 | gpte = lgread(cpu, gpte_addr(cpu, gpmd, vaddr), pte_t); |
| 652 | #else |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 653 | gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t); |
Rusty Russell | 92b4d8d | 2009-06-12 22:27:08 -0600 | [diff] [blame] | 654 | #endif |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 655 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 656 | kill_guest(cpu, "Bad address %#lx", vaddr); |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 657 | |
| 658 | return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK); |
| 659 | } |
| 660 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 661 | /* |
| 662 | * We keep several page tables. This is a simple routine to find the page |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 663 | * table (if any) corresponding to this top-level address the Guest has given |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 664 | * us. |
| 665 | */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 666 | static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) |
| 667 | { |
| 668 | unsigned int i; |
| 669 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
Rusty Russell | 4357bd9 | 2008-03-11 09:35:57 -0500 | [diff] [blame] | 670 | if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 671 | break; |
| 672 | return i; |
| 673 | } |
| 674 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 675 | /*H:435 |
| 676 | * And this is us, creating the new page directory. If we really do |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 677 | * allocate a new one (and so the kernel parts are not there), we set |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 678 | * blank_pgdir. |
| 679 | */ |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 680 | static unsigned int new_pgdir(struct lg_cpu *cpu, |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 681 | unsigned long gpgdir, |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 682 | int *blank_pgdir) |
| 683 | { |
| 684 | unsigned int next; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 685 | #ifdef CONFIG_X86_PAE |
| 686 | pmd_t *pmd_table; |
| 687 | #endif |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 688 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 689 | /* |
| 690 | * We pick one entry at random to throw out. Choosing the Least |
| 691 | * Recently Used might be better, but this is easy. |
| 692 | */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 693 | next = random32() % ARRAY_SIZE(cpu->lg->pgdirs); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 694 | /* If it's never been allocated at all before, try now. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 695 | if (!cpu->lg->pgdirs[next].pgdir) { |
| 696 | cpu->lg->pgdirs[next].pgdir = |
| 697 | (pgd_t *)get_zeroed_page(GFP_KERNEL); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 698 | /* If the allocation fails, just keep using the one we have */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 699 | if (!cpu->lg->pgdirs[next].pgdir) |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 700 | next = cpu->cpu_pgd; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 701 | else { |
| 702 | #ifdef CONFIG_X86_PAE |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 703 | /* |
| 704 | * In PAE mode, allocate a pmd page and populate the |
| 705 | * last pgd entry. |
| 706 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 707 | pmd_table = (pmd_t *)get_zeroed_page(GFP_KERNEL); |
| 708 | if (!pmd_table) { |
| 709 | free_page((long)cpu->lg->pgdirs[next].pgdir); |
| 710 | set_pgd(cpu->lg->pgdirs[next].pgdir, __pgd(0)); |
| 711 | next = cpu->cpu_pgd; |
| 712 | } else { |
| 713 | set_pgd(cpu->lg->pgdirs[next].pgdir + |
| 714 | SWITCHER_PGD_INDEX, |
| 715 | __pgd(__pa(pmd_table) | _PAGE_PRESENT)); |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 716 | /* |
| 717 | * This is a blank page, so there are no kernel |
| 718 | * mappings: caller must map the stack! |
| 719 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 720 | *blank_pgdir = 1; |
| 721 | } |
| 722 | #else |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 723 | *blank_pgdir = 1; |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 724 | #endif |
| 725 | } |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 726 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 727 | /* Record which Guest toplevel this shadows. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 728 | cpu->lg->pgdirs[next].gpgdir = gpgdir; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 729 | /* Release all the non-kernel mappings. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 730 | flush_user_mappings(cpu->lg, next); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 731 | |
| 732 | return next; |
| 733 | } |
| 734 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 735 | /*H:470 |
| 736 | * Finally, a routine which throws away everything: all PGD entries in all |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 737 | * the shadow page tables, including the Guest's kernel mappings. This is used |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 738 | * when we destroy the Guest. |
| 739 | */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 740 | static void release_all_pagetables(struct lguest *lg) |
| 741 | { |
| 742 | unsigned int i, j; |
| 743 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 744 | /* Every shadow pagetable this Guest has */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 745 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 746 | if (lg->pgdirs[i].pgdir) { |
| 747 | #ifdef CONFIG_X86_PAE |
| 748 | pgd_t *spgd; |
| 749 | pmd_t *pmdpage; |
| 750 | unsigned int k; |
| 751 | |
| 752 | /* Get the last pmd page. */ |
| 753 | spgd = lg->pgdirs[i].pgdir + SWITCHER_PGD_INDEX; |
| 754 | pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); |
| 755 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 756 | /* |
| 757 | * And release the pmd entries of that pmd page, |
| 758 | * except for the switcher pmd. |
| 759 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 760 | for (k = 0; k < SWITCHER_PMD_INDEX; k++) |
| 761 | release_pmd(&pmdpage[k]); |
| 762 | #endif |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 763 | /* Every PGD entry except the Switcher at the top */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 764 | for (j = 0; j < SWITCHER_PGD_INDEX; j++) |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 765 | release_pgd(lg->pgdirs[i].pgdir + j); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 766 | } |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 767 | } |
| 768 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 769 | /* |
| 770 | * We also throw away everything when a Guest tells us it's changed a kernel |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 771 | * 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] | 772 | * throw them all away. This traps the Guest in amber for a while as |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 773 | * everything faults back in, but it's rare. |
| 774 | */ |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 775 | void guest_pagetable_clear_all(struct lg_cpu *cpu) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 776 | { |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 777 | release_all_pagetables(cpu->lg); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 778 | /* We need the Guest kernel stack mapped again. */ |
Glauber de Oliveira Costa | 4665ac8e | 2008-01-07 11:05:35 -0200 | [diff] [blame] | 779 | pin_stack_pages(cpu); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 780 | } |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 781 | |
| 782 | /*H:430 |
| 783 | * (iv) Switching page tables |
| 784 | * |
| 785 | * Now we've seen all the page table setting and manipulation, let's see |
| 786 | * what happens when the Guest changes page tables (ie. changes the top-level |
| 787 | * pgdir). This occurs on almost every context switch. |
| 788 | */ |
| 789 | void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable) |
| 790 | { |
| 791 | int newpgdir, repin = 0; |
| 792 | |
| 793 | /* |
| 794 | * The very first time they call this, we're actually running without |
| 795 | * any page tables; we've been making it up. Throw them away now. |
| 796 | */ |
| 797 | if (unlikely(cpu->linear_pages)) { |
| 798 | release_all_pagetables(cpu->lg); |
| 799 | cpu->linear_pages = false; |
| 800 | /* Force allocation of a new pgdir. */ |
| 801 | newpgdir = ARRAY_SIZE(cpu->lg->pgdirs); |
| 802 | } else { |
| 803 | /* Look to see if we have this one already. */ |
| 804 | newpgdir = find_pgdir(cpu->lg, pgtable); |
| 805 | } |
| 806 | |
| 807 | /* |
| 808 | * If not, we allocate or mug an existing one: if it's a fresh one, |
| 809 | * repin gets set to 1. |
| 810 | */ |
| 811 | if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs)) |
| 812 | newpgdir = new_pgdir(cpu, pgtable, &repin); |
| 813 | /* Change the current pgd index to the new one. */ |
| 814 | cpu->cpu_pgd = newpgdir; |
| 815 | /* If it was completely blank, we map in the Guest kernel stack */ |
| 816 | if (repin) |
| 817 | pin_stack_pages(cpu); |
| 818 | } |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 819 | /*:*/ |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 820 | |
| 821 | /*M:009 |
| 822 | * Since we throw away all mappings when a kernel mapping changes, our |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 823 | * performance sucks for guests using highmem. In fact, a guest with |
| 824 | * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is |
| 825 | * usually slower than a Guest with less memory. |
| 826 | * |
| 827 | * This, of course, cannot be fixed. It would take some kind of... well, I |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 828 | * don't know, but the term "puissant code-fu" comes to mind. |
| 829 | :*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 830 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 831 | /*H:420 |
| 832 | * This is the routine which actually sets the page table entry for then |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 833 | * "idx"'th shadow page table. |
| 834 | * |
| 835 | * Normally, we can just throw out the old entry and replace it with 0: if they |
| 836 | * use it demand_page() will put the new entry in. We need to do this anyway: |
| 837 | * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page |
| 838 | * is read from, and _PAGE_DIRTY when it's written to. |
| 839 | * |
| 840 | * But Avi Kivity pointed out that most Operating Systems (Linux included) set |
| 841 | * these bits on PTEs immediately anyway. This is done to save the CPU from |
| 842 | * having to update them, but it helps us the same way: if they set |
| 843 | * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if |
| 844 | * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately. |
| 845 | */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 846 | static void do_set_pte(struct lg_cpu *cpu, int idx, |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 847 | unsigned long vaddr, pte_t gpte) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 848 | { |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 849 | /* Look up the matching shadow page directory entry. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 850 | pgd_t *spgd = spgd_addr(cpu, idx, vaddr); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 851 | #ifdef CONFIG_X86_PAE |
| 852 | pmd_t *spmd; |
| 853 | #endif |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 854 | |
| 855 | /* 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] | 856 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 857 | #ifdef CONFIG_X86_PAE |
| 858 | spmd = spmd_addr(cpu, *spgd, vaddr); |
| 859 | if (pmd_flags(*spmd) & _PAGE_PRESENT) { |
| 860 | #endif |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 861 | /* Otherwise, start by releasing the existing entry. */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 862 | pte_t *spte = spte_addr(cpu, *spgd, vaddr); |
| 863 | release_pte(*spte); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 864 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 865 | /* |
| 866 | * If they're setting this entry as dirty or accessed, |
| 867 | * we might as well put that entry they've given us in |
| 868 | * now. This shaves 10% off a copy-on-write |
| 869 | * micro-benchmark. |
| 870 | */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 871 | if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) { |
Rusty Russell | e1d1260 | 2013-04-22 14:10:39 +0930 | [diff] [blame^] | 872 | if (!check_gpte(cpu, gpte)) |
| 873 | return; |
Rusty Russell | 4c1ea3d | 2009-09-23 22:26:45 -0600 | [diff] [blame] | 874 | set_pte(spte, |
| 875 | gpte_to_spte(cpu, gpte, |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 876 | pte_flags(gpte) & _PAGE_DIRTY)); |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 877 | } else { |
| 878 | /* |
| 879 | * Otherwise kill it and we can demand_page() |
| 880 | * it in later. |
| 881 | */ |
Rusty Russell | 4c1ea3d | 2009-09-23 22:26:45 -0600 | [diff] [blame] | 882 | set_pte(spte, __pte(0)); |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 883 | } |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 884 | #ifdef CONFIG_X86_PAE |
| 885 | } |
| 886 | #endif |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 887 | } |
| 888 | } |
| 889 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 890 | /*H:410 |
| 891 | * Updating a PTE entry is a little trickier. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 892 | * |
| 893 | * We keep track of several different page tables (the Guest uses one for each |
| 894 | * process, so it makes sense to cache at least a few). Each of these have |
| 895 | * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for |
| 896 | * all processes. So when the page table above that address changes, we update |
| 897 | * all the page tables, not just the current one. This is rare. |
| 898 | * |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 899 | * The benefit is that when we have to track a new page table, we can keep all |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 900 | * the kernel mappings. This speeds up context switch immensely. |
| 901 | */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 902 | void guest_set_pte(struct lg_cpu *cpu, |
Rusty Russell | ee3db0f | 2007-10-22 11:03:34 +1000 | [diff] [blame] | 903 | unsigned long gpgdir, unsigned long vaddr, pte_t gpte) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 904 | { |
Rusty Russell | 68a644d | 2013-04-22 14:10:37 +0930 | [diff] [blame] | 905 | /* We don't let you remap the Switcher; we need it to get back! */ |
| 906 | if (vaddr >= switcher_addr) { |
| 907 | kill_guest(cpu, "attempt to set pte into Switcher pages"); |
| 908 | return; |
| 909 | } |
| 910 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 911 | /* |
| 912 | * Kernel mappings must be changed on all top levels. Slow, but doesn't |
| 913 | * happen often. |
| 914 | */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 915 | if (vaddr >= cpu->lg->kernel_address) { |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 916 | unsigned int i; |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 917 | for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++) |
| 918 | if (cpu->lg->pgdirs[i].pgdir) |
| 919 | do_set_pte(cpu, i, vaddr, gpte); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 920 | } else { |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 921 | /* Is this page table one we have a shadow for? */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 922 | int pgdir = find_pgdir(cpu->lg, gpgdir); |
| 923 | if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs)) |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 924 | /* If so, do the update. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 925 | do_set_pte(cpu, pgdir, vaddr, gpte); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 926 | } |
| 927 | } |
| 928 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 929 | /*H:400 |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 930 | * (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] | 931 | * |
| 932 | * Just like we did in interrupts_and_traps.c, it makes sense for us to deal |
| 933 | * with the other side of page tables while we're here: what happens when the |
| 934 | * Guest asks for a page table to be updated? |
| 935 | * |
| 936 | * We already saw that demand_page() will fill in the shadow page tables when |
| 937 | * needed, so we can simply remove shadow page table entries whenever the Guest |
| 938 | * tells us they've changed. When the Guest tries to use the new entry it will |
| 939 | * fault and demand_page() will fix it up. |
| 940 | * |
Anand Gadiyar | fd589a8 | 2009-07-16 17:13:03 +0200 | [diff] [blame] | 941 | * So with that in mind here's our code to update a (top-level) PGD entry: |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 942 | */ |
Matias Zabaljauregui | ebe0ba8 | 2009-05-30 15:48:08 -0300 | [diff] [blame] | 943 | void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 944 | { |
| 945 | int pgdir; |
| 946 | |
| 947 | if (idx >= SWITCHER_PGD_INDEX) |
| 948 | return; |
| 949 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 950 | /* 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] | 951 | pgdir = find_pgdir(lg, gpgdir); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 952 | if (pgdir < ARRAY_SIZE(lg->pgdirs)) |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 953 | /* ... throw it away. */ |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 954 | release_pgd(lg->pgdirs[pgdir].pgdir + idx); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 955 | } |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 956 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 957 | #ifdef CONFIG_X86_PAE |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 958 | /* For setting a mid-level, we just throw everything away. It's easy. */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 959 | void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx) |
| 960 | { |
| 961 | guest_pagetable_clear_all(&lg->cpus[0]); |
| 962 | } |
| 963 | #endif |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 964 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 965 | /*H:500 |
| 966 | * (vii) Setting up the page tables initially. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 967 | * |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 968 | * When a Guest is first created, set initialize a shadow page table which |
| 969 | * we will populate on future faults. The Guest doesn't have any actual |
| 970 | * pagetables yet, so we set linear_pages to tell demand_page() to fake it |
| 971 | * for the moment. |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 972 | */ |
Matias Zabaljauregui | 58a2456 | 2008-09-29 01:40:07 -0300 | [diff] [blame] | 973 | int init_guest_pagetable(struct lguest *lg) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 974 | { |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 975 | struct lg_cpu *cpu = &lg->cpus[0]; |
| 976 | int allocated = 0; |
Matias Zabaljauregui | 58a2456 | 2008-09-29 01:40:07 -0300 | [diff] [blame] | 977 | |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 978 | /* lg (and lg->cpus[]) starts zeroed: this allocates a new pgdir */ |
| 979 | cpu->cpu_pgd = new_pgdir(cpu, 0, &allocated); |
| 980 | if (!allocated) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 981 | return -ENOMEM; |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 982 | |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 983 | /* We start with a linear mapping until the initialize. */ |
| 984 | cpu->linear_pages = true; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 985 | return 0; |
| 986 | } |
| 987 | |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 988 | /*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 989 | void page_table_guest_data_init(struct lg_cpu *cpu) |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 990 | { |
Rusty Russell | c215a8b | 2013-04-22 14:10:37 +0930 | [diff] [blame] | 991 | /* |
| 992 | * We tell the Guest that it can't use the virtual addresses |
| 993 | * used by the Switcher. This trick is equivalent to 4GB - |
| 994 | * switcher_addr. |
| 995 | */ |
| 996 | u32 top = ~switcher_addr + 1; |
| 997 | |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 998 | /* We get the kernel address: above this is all kernel memory. */ |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 999 | if (get_user(cpu->lg->kernel_address, |
Rusty Russell | c215a8b | 2013-04-22 14:10:37 +0930 | [diff] [blame] | 1000 | &cpu->lg->lguest_data->kernel_address) |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1001 | /* |
Rusty Russell | c215a8b | 2013-04-22 14:10:37 +0930 | [diff] [blame] | 1002 | * We tell the Guest that it can't use the top virtual |
| 1003 | * addresses (used by the Switcher). |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1004 | */ |
Rusty Russell | c215a8b | 2013-04-22 14:10:37 +0930 | [diff] [blame] | 1005 | || put_user(top, &cpu->lg->lguest_data->reserve_mem)) { |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 1006 | kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data); |
Rusty Russell | 5dea1c8 | 2011-07-22 14:39:48 +0930 | [diff] [blame] | 1007 | return; |
| 1008 | } |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 1009 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1010 | /* |
| 1011 | * In flush_user_mappings() we loop from 0 to |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 1012 | * "pgd_index(lg->kernel_address)". This assumes it won't hit the |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1013 | * Switcher mappings, so check that now. |
| 1014 | */ |
Rusty Russell | 68a644d | 2013-04-22 14:10:37 +0930 | [diff] [blame] | 1015 | if (cpu->lg->kernel_address >= switcher_addr) |
Glauber de Oliveira Costa | 382ac6b | 2008-01-17 19:19:42 -0200 | [diff] [blame] | 1016 | kill_guest(cpu, "bad kernel address %#lx", |
| 1017 | cpu->lg->kernel_address); |
Rusty Russell | 47436aa | 2007-10-22 11:03:36 +1000 | [diff] [blame] | 1018 | } |
| 1019 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1020 | /* When a Guest dies, our cleanup is fairly simple. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1021 | void free_guest_pagetable(struct lguest *lg) |
| 1022 | { |
| 1023 | unsigned int i; |
| 1024 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1025 | /* Throw away all page table pages. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1026 | release_all_pagetables(lg); |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1027 | /* Now free the top levels: free_page() can handle 0 just fine. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1028 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
| 1029 | free_page((long)lg->pgdirs[i].pgdir); |
| 1030 | } |
| 1031 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1032 | /*H:480 |
| 1033 | * (vi) Mapping the Switcher when the Guest is about to run. |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1034 | * |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 1035 | * 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] | 1036 | * 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] | 1037 | * for each CPU already set up, we just need to hook them in now we know which |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1038 | * Guest is about to run on this CPU. |
| 1039 | */ |
Glauber de Oliveira Costa | 0c78441 | 2008-01-07 11:05:30 -0200 | [diff] [blame] | 1040 | 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] | 1041 | { |
Christoph Lameter | c9f2954 | 2010-11-30 13:07:21 -0600 | [diff] [blame] | 1042 | pte_t *switcher_pte_page = __this_cpu_read(switcher_pte_pages); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 1043 | pte_t regs_pte; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1044 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 1045 | #ifdef CONFIG_X86_PAE |
| 1046 | pmd_t switcher_pmd; |
| 1047 | pmd_t *pmd_table; |
| 1048 | |
Rusty Russell | 4c1ea3d | 2009-09-23 22:26:45 -0600 | [diff] [blame] | 1049 | switcher_pmd = pfn_pmd(__pa(switcher_pte_page) >> PAGE_SHIFT, |
| 1050 | PAGE_KERNEL_EXEC); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 1051 | |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1052 | /* Figure out where the pmd page is, by reading the PGD, and converting |
| 1053 | * it to a virtual address. */ |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 1054 | pmd_table = __va(pgd_pfn(cpu->lg-> |
| 1055 | pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX]) |
| 1056 | << PAGE_SHIFT); |
Rusty Russell | a91d74a | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1057 | /* Now write it into the shadow page table. */ |
Rusty Russell | 4c1ea3d | 2009-09-23 22:26:45 -0600 | [diff] [blame] | 1058 | set_pmd(&pmd_table[SWITCHER_PMD_INDEX], switcher_pmd); |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 1059 | #else |
| 1060 | pgd_t switcher_pgd; |
| 1061 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1062 | /* |
| 1063 | * Make the last PGD entry for this Guest point to the Switcher's PTE |
| 1064 | * page for this CPU (with appropriate flags). |
| 1065 | */ |
Matias Zabaljauregui | ed1dc77 | 2009-05-30 15:35:49 -0300 | [diff] [blame] | 1066 | switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL_EXEC); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 1067 | |
Glauber de Oliveira Costa | 1713608 | 2008-01-07 11:05:37 -0200 | [diff] [blame] | 1068 | cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1069 | |
Matias Zabaljauregui | acdd0b6 | 2009-06-12 22:27:07 -0600 | [diff] [blame] | 1070 | #endif |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1071 | /* |
| 1072 | * We also change the Switcher PTE page. When we're running the Guest, |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1073 | * we want the Guest's "regs" page to appear where the first Switcher |
| 1074 | * page for this CPU is. This is an optimization: when the Switcher |
| 1075 | * saves the Guest registers, it saves them into the first page of this |
| 1076 | * CPU's "struct lguest_pages": if we make sure the Guest's register |
| 1077 | * page is already mapped there, we don't have to copy them out |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1078 | * again. |
| 1079 | */ |
Rusty Russell | 4c1ea3d | 2009-09-23 22:26:45 -0600 | [diff] [blame] | 1080 | regs_pte = pfn_pte(__pa(cpu->regs_page) >> PAGE_SHIFT, PAGE_KERNEL); |
| 1081 | set_pte(&switcher_pte_page[pte_index((unsigned long)pages)], regs_pte); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1082 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1083 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1084 | |
| 1085 | static void free_switcher_pte_pages(void) |
| 1086 | { |
| 1087 | unsigned int i; |
| 1088 | |
| 1089 | for_each_possible_cpu(i) |
| 1090 | free_page((long)switcher_pte_page(i)); |
| 1091 | } |
| 1092 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1093 | /*H:520 |
| 1094 | * Setting up the Switcher PTE page for given CPU is fairly easy, given |
Rusty Russell | 93a2cdf | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1095 | * the CPU number and the "struct page"s for the Switcher and per-cpu pages. |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1096 | */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1097 | static __init void populate_switcher_pte_page(unsigned int cpu, |
Rusty Russell | 93a2cdf | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1098 | struct page *switcher_pages[]) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1099 | { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 1100 | pte_t *pte = switcher_pte_page(cpu); |
Rusty Russell | 93a2cdf | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1101 | int i; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1102 | |
Rusty Russell | 93a2cdf | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1103 | /* The first entries maps the Switcher code. */ |
| 1104 | set_pte(&pte[0], mk_pte(switcher_pages[0], |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 1105 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED))); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1106 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1107 | /* The only other thing we map is this CPU's pair of pages. */ |
Rusty Russell | 93a2cdf | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1108 | i = 1 + cpu*2; |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1109 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1110 | /* First page (Guest registers) is writable from the Guest */ |
Rusty Russell | 856c608 | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1111 | set_pte(&pte[i], pfn_pte(page_to_pfn(switcher_pages[i]), |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 1112 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW))); |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 1113 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1114 | /* |
| 1115 | * The second page contains the "struct lguest_ro_state", and is |
| 1116 | * read-only. |
| 1117 | */ |
Rusty Russell | 856c608 | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1118 | set_pte(&pte[i+1], pfn_pte(page_to_pfn(switcher_pages[i+1]), |
Matias Zabaljauregui | 90603d1 | 2009-06-12 22:27:06 -0600 | [diff] [blame] | 1119 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED))); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1120 | } |
| 1121 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1122 | /* |
| 1123 | * We've made it through the page table code. Perhaps our tired brains are |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 1124 | * still processing the details, or perhaps we're simply glad it's over. |
| 1125 | * |
Rusty Russell | a6bd8e1 | 2008-03-28 11:05:53 -0500 | [diff] [blame] | 1126 | * If nothing else, note that all this complexity in juggling shadow page tables |
| 1127 | * in sync with the Guest's page tables is for one reason: for most Guests this |
| 1128 | * page table dance determines how bad performance will be. This is why Xen |
| 1129 | * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD |
| 1130 | * have implemented shadow page table support directly into hardware. |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 1131 | * |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1132 | * There is just one file remaining in the Host. |
| 1133 | */ |
Rusty Russell | e1e7296 | 2007-10-25 15:02:50 +1000 | [diff] [blame] | 1134 | |
Rusty Russell | 2e04ef7 | 2009-07-30 16:03:45 -0600 | [diff] [blame] | 1135 | /*H:510 |
| 1136 | * At boot or module load time, init_pagetables() allocates and populates |
| 1137 | * the Switcher PTE page for each CPU. |
| 1138 | */ |
Rusty Russell | 93a2cdf | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1139 | __init int init_pagetables(struct page **switcher_pages) |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1140 | { |
| 1141 | unsigned int i; |
| 1142 | |
| 1143 | for_each_possible_cpu(i) { |
Matias Zabaljauregui | df29f43 | 2007-10-22 11:03:33 +1000 | [diff] [blame] | 1144 | switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1145 | if (!switcher_pte_page(i)) { |
| 1146 | free_switcher_pte_pages(); |
| 1147 | return -ENOMEM; |
| 1148 | } |
Rusty Russell | 93a2cdf | 2013-04-22 14:10:38 +0930 | [diff] [blame] | 1149 | populate_switcher_pte_page(i, switcher_pages); |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1150 | } |
| 1151 | return 0; |
| 1152 | } |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1153 | /*:*/ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1154 | |
Rusty Russell | bff672e | 2007-07-26 10:41:04 -0700 | [diff] [blame] | 1155 | /* Cleaning up simply involves freeing the PTE page for each CPU. */ |
Rusty Russell | d7e28ff | 2007-07-19 01:49:23 -0700 | [diff] [blame] | 1156 | void free_pagetables(void) |
| 1157 | { |
| 1158 | free_switcher_pte_pages(); |
| 1159 | } |