diff options
Diffstat (limited to 'drivers/lguest/page_tables.c')
| -rw-r--r-- | drivers/lguest/page_tables.c | 1086 |
1 files changed, 770 insertions, 316 deletions
diff --git a/drivers/lguest/page_tables.c b/drivers/lguest/page_tables.c index d93500f24fb..e8b55c3a617 100644 --- a/drivers/lguest/page_tables.c +++ b/drivers/lguest/page_tables.c @@ -1,13 +1,16 @@ -/*P:700 The pagetable code, on the other hand, still shows the scars of +/*P:700 + * The pagetable code, on the other hand, still shows the scars of * previous encounters. It's functional, and as neat as it can be in the * circumstances, but be wary, for these things are subtle and break easily. * The Guest provides a virtual to physical mapping, but we can neither trust * it nor use it: we verify and convert it here then point the CPU to the - * converted Guest pages when running the Guest. :*/ + * converted Guest pages when running the Guest. +:*/ -/* Copyright (C) Rusty Russell IBM Corporation 2006. +/* Copyright (C) Rusty Russell IBM Corporation 2013. * GPL v2 and any later version */ #include <linux/mm.h> +#include <linux/gfp.h> #include <linux/types.h> #include <linux/spinlock.h> #include <linux/random.h> @@ -16,18 +19,20 @@ #include <asm/uaccess.h> #include "lg.h" -/*M:008 We hold reference to pages, which prevents them from being swapped. +/*M:008 + * We hold reference to pages, which prevents them from being swapped. * It'd be nice to have a callback in the "struct mm_struct" when Linux wants * to swap out. If we had this, and a shrinker callback to trim PTE pages, we - * could probably consider launching Guests as non-root. :*/ + * could probably consider launching Guests as non-root. +:*/ /*H:300 * The Page Table Code * - * We use two-level page tables for the Guest. If you're not entirely - * comfortable with virtual addresses, physical addresses and page tables then - * I recommend you review arch/x86/lguest/boot.c's "Page Table Handling" (with - * diagrams!). + * We use two-level page tables for the Guest, or three-level with PAE. If + * you're not entirely comfortable with virtual addresses, physical addresses + * and page tables then I recommend you review arch/x86/lguest/boot.c's "Page + * Table Handling" (with diagrams!). * * The Guest keeps page tables, but we maintain the actual ones here: these are * called "shadow" page tables. Which is a very Guest-centric name: these are @@ -44,22 +49,28 @@ * (v) Flushing (throwing away) page tables, * (vi) Mapping the Switcher when the Guest is about to run, * (vii) Setting up the page tables initially. - :*/ +:*/ - -/* 1024 entries in a page table page maps 1024 pages: 4MB. The Switcher is - * conveniently placed at the top 4MB, so it uses a separate, complete PTE - * page. */ +/* + * The Switcher uses the complete top PTE page. That's 1024 PTE entries (4MB) + * or 512 PTE entries with PAE (2MB). + */ #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1) -/* We actually need a separate PTE page for each CPU. Remember that after the - * Switcher code itself comes two pages for each CPU, and we don't want this - * CPU's guest to see the pages of any other CPU. */ -static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); -#define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) - -/*H:320 The page table code is curly enough to need helper functions to keep it - * clear and clean. +/* + * For PAE we need the PMD index as well. We use the last 2MB, so we + * will need the last pmd entry of the last pmd page. + */ +#ifdef CONFIG_X86_PAE +#define CHECK_GPGD_MASK _PAGE_PRESENT +#else +#define CHECK_GPGD_MASK _PAGE_TABLE +#endif + +/*H:320 + * The page table code is curly enough to need helper functions to keep it + * clear and clean. The kernel itself provides many of them; one advantage + * of insisting that the Guest and Host use the same CONFIG_X86_PAE setting. * * There are two functions which return pointers to the shadow (aka "real") * page tables. @@ -67,104 +78,160 @@ static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); * spgd_addr() takes the virtual address and returns a pointer to the top-level * page directory entry (PGD) for that address. Since we keep track of several * page tables, the "i" argument tells us which one we're interested in (it's - * usually the current one). */ + * usually the current one). + */ static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr) { unsigned int index = pgd_index(vaddr); - /* We kill any Guest trying to touch the Switcher addresses. */ - if (index >= SWITCHER_PGD_INDEX) { - kill_guest(cpu, "attempt to access switcher pages"); - index = 0; - } /* Return a pointer index'th pgd entry for the i'th page table. */ return &cpu->lg->pgdirs[i].pgdir[index]; } -/* This routine then takes the page directory entry returned above, which +#ifdef CONFIG_X86_PAE +/* + * This routine then takes the PGD entry given above, which contains the + * address of the PMD page. It then returns a pointer to the PMD entry for the + * given address. + */ +static pmd_t *spmd_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) +{ + unsigned int index = pmd_index(vaddr); + pmd_t *page; + + /* You should never call this if the PGD entry wasn't valid */ + BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); + page = __va(pgd_pfn(spgd) << PAGE_SHIFT); + + return &page[index]; +} +#endif + +/* + * This routine then takes the page directory entry returned above, which * contains the address of the page table entry (PTE) page. It then returns a - * pointer to the PTE entry for the given address. */ -static pte_t *spte_addr(pgd_t spgd, unsigned long vaddr) + * pointer to the PTE entry for the given address. + */ +static pte_t *spte_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) { +#ifdef CONFIG_X86_PAE + pmd_t *pmd = spmd_addr(cpu, spgd, vaddr); + pte_t *page = __va(pmd_pfn(*pmd) << PAGE_SHIFT); + + /* You should never call this if the PMD entry wasn't valid */ + BUG_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)); +#else pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT); /* You should never call this if the PGD entry wasn't valid */ BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); - return &page[(vaddr >> PAGE_SHIFT) % PTRS_PER_PTE]; +#endif + + return &page[pte_index(vaddr)]; } -/* These two functions just like the above two, except they access the Guest - * page tables. Hence they return a Guest address. */ +/* + * These functions are just like the above, except they access the Guest + * page tables. Hence they return a Guest address. + */ static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr) { unsigned int index = vaddr >> (PGDIR_SHIFT); return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t); } -static unsigned long gpte_addr(pgd_t gpgd, unsigned long vaddr) +#ifdef CONFIG_X86_PAE +/* Follow the PGD to the PMD. */ +static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr) { unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); - return gpage + ((vaddr>>PAGE_SHIFT) % PTRS_PER_PTE) * sizeof(pte_t); + return gpage + pmd_index(vaddr) * sizeof(pmd_t); +} + +/* Follow the PMD to the PTE. */ +static unsigned long gpte_addr(struct lg_cpu *cpu, + pmd_t gpmd, unsigned long vaddr) +{ + unsigned long gpage = pmd_pfn(gpmd) << PAGE_SHIFT; + + BUG_ON(!(pmd_flags(gpmd) & _PAGE_PRESENT)); + return gpage + pte_index(vaddr) * sizeof(pte_t); +} +#else +/* Follow the PGD to the PTE (no mid-level for !PAE). */ +static unsigned long gpte_addr(struct lg_cpu *cpu, + pgd_t gpgd, unsigned long vaddr) +{ + unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; + + BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); + return gpage + pte_index(vaddr) * sizeof(pte_t); } +#endif /*:*/ -/*M:014 get_pfn is slow; it takes the mmap sem and calls get_user_pages. We - * could probably try to grab batches of pages here as an optimization - * (ie. pre-faulting). :*/ +/*M:007 + * get_pfn is slow: we could probably try to grab batches of pages here as + * an optimization (ie. pre-faulting). +:*/ -/*H:350 This routine takes a page number given by the Guest and converts it to +/*H:350 + * This routine takes a page number given by the Guest and converts it to * an actual, physical page number. It can fail for several reasons: the * virtual address might not be mapped by the Launcher, the write flag is set * and the page is read-only, or the write flag was set and the page was * shared so had to be copied, but we ran out of memory. * * This holds a reference to the page, so release_pte() is careful to put that - * back. */ + * back. + */ static unsigned long get_pfn(unsigned long virtpfn, int write) { struct page *page; - /* This value indicates failure. */ - unsigned long ret = -1UL; - /* get_user_pages() is a complex interface: it gets the "struct - * vm_area_struct" and "struct page" assocated with a range of pages. - * It also needs the task's mmap_sem held, and is not very quick. - * It returns the number of pages it got. */ - down_read(¤t->mm->mmap_sem); - if (get_user_pages(current, current->mm, virtpfn << PAGE_SHIFT, - 1, write, 1, &page, NULL) == 1) - ret = page_to_pfn(page); - up_read(¤t->mm->mmap_sem); - return ret; + /* gup me one page at this address please! */ + if (get_user_pages_fast(virtpfn << PAGE_SHIFT, 1, write, &page) == 1) + return page_to_pfn(page); + + /* This value indicates failure. */ + return -1UL; } -/*H:340 Converting a Guest page table entry to a shadow (ie. real) page table +/*H:340 + * Converting a Guest page table entry to a shadow (ie. real) page table * entry can be a little tricky. The flags are (almost) the same, but the * Guest PTE contains a virtual page number: the CPU needs the real page - * number. */ + * number. + */ static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write) { unsigned long pfn, base, flags; - /* The Guest sets the global flag, because it thinks that it is using + /* + * The Guest sets the global flag, because it thinks that it is using * PGE. We only told it to use PGE so it would tell us whether it was * flushing a kernel mapping or a userspace mapping. We don't actually - * use the global bit, so throw it away. */ + * use the global bit, so throw it away. + */ flags = (pte_flags(gpte) & ~_PAGE_GLOBAL); /* The Guest's pages are offset inside the Launcher. */ base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE; - /* We need a temporary "unsigned long" variable to hold the answer from + /* + * We need a temporary "unsigned long" variable to hold the answer from * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't * fit in spte.pfn. get_pfn() finds the real physical number of the - * page, given the virtual number. */ + * page, given the virtual number. + */ pfn = get_pfn(base + pte_pfn(gpte), write); if (pfn == -1UL) { kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte)); - /* When we destroy the Guest, we'll go through the shadow page + /* + * When we destroy the Guest, we'll go through the shadow page * tables and release_pte() them. Make sure we don't think - * this one is valid! */ + * this one is valid! + */ flags = 0; } /* Now we assemble our shadow PTE from the page number and flags. */ @@ -174,25 +241,127 @@ static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write) /*H:460 And to complete the chain, release_pte() looks like this: */ static void release_pte(pte_t pte) { - /* Remember that get_user_pages() took a reference to the page, in - * get_pfn()? We have to put it back now. */ + /* + * Remember that get_user_pages_fast() took a reference to the page, in + * get_pfn()? We have to put it back now. + */ if (pte_flags(pte) & _PAGE_PRESENT) - put_page(pfn_to_page(pte_pfn(pte))); + put_page(pte_page(pte)); } /*:*/ -static void check_gpte(struct lg_cpu *cpu, pte_t gpte) +static bool check_gpte(struct lg_cpu *cpu, pte_t gpte) { if ((pte_flags(gpte) & _PAGE_PSE) || - pte_pfn(gpte) >= cpu->lg->pfn_limit) + pte_pfn(gpte) >= cpu->lg->pfn_limit) { kill_guest(cpu, "bad page table entry"); + return false; + } + return true; } -static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd) +static bool check_gpgd(struct lg_cpu *cpu, pgd_t gpgd) { - if ((pgd_flags(gpgd) & ~_PAGE_TABLE) || - (pgd_pfn(gpgd) >= cpu->lg->pfn_limit)) + if ((pgd_flags(gpgd) & ~CHECK_GPGD_MASK) || + (pgd_pfn(gpgd) >= cpu->lg->pfn_limit)) { kill_guest(cpu, "bad page directory entry"); + return false; + } + return true; +} + +#ifdef CONFIG_X86_PAE +static bool check_gpmd(struct lg_cpu *cpu, pmd_t gpmd) +{ + if ((pmd_flags(gpmd) & ~_PAGE_TABLE) || + (pmd_pfn(gpmd) >= cpu->lg->pfn_limit)) { + kill_guest(cpu, "bad page middle directory entry"); + return false; + } + return true; +} +#endif + +/*H:331 + * This is the core routine to walk the shadow page tables and find the page + * table entry for a specific address. + * + * If allocate is set, then we allocate any missing levels, setting the flags + * on the new page directory and mid-level directories using the arguments + * (which are copied from the Guest's page table entries). + */ +static pte_t *find_spte(struct lg_cpu *cpu, unsigned long vaddr, bool allocate, + int pgd_flags, int pmd_flags) +{ + pgd_t *spgd; + /* Mid level for PAE. */ +#ifdef CONFIG_X86_PAE + pmd_t *spmd; +#endif + + /* Get top level entry. */ + spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); + if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) { + /* No shadow entry: allocate a new shadow PTE page. */ + unsigned long ptepage; + + /* If they didn't want us to allocate anything, stop. */ + if (!allocate) + return NULL; + + ptepage = get_zeroed_page(GFP_KERNEL); + /* + * This is not really the Guest's fault, but killing it is + * simple for this corner case. + */ + if (!ptepage) { + kill_guest(cpu, "out of memory allocating pte page"); + return NULL; + } + /* + * And we copy the flags to the shadow PGD entry. The page + * number in the shadow PGD is the page we just allocated. + */ + set_pgd(spgd, __pgd(__pa(ptepage) | pgd_flags)); + } + + /* + * Intel's Physical Address Extension actually uses three levels of + * page tables, so we need to look in the mid-level. + */ +#ifdef CONFIG_X86_PAE + /* Now look at the mid-level shadow entry. */ + spmd = spmd_addr(cpu, *spgd, vaddr); + + if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) { + /* No shadow entry: allocate a new shadow PTE page. */ + unsigned long ptepage; + + /* If they didn't want us to allocate anything, stop. */ + if (!allocate) + return NULL; + + ptepage = get_zeroed_page(GFP_KERNEL); + + /* + * This is not really the Guest's fault, but killing it is + * simple for this corner case. + */ + if (!ptepage) { + kill_guest(cpu, "out of memory allocating pmd page"); + return NULL; + } + + /* + * And we copy the flags to the shadow PMD entry. The page + * number in the shadow PMD is the page we just allocated. + */ + set_pmd(spmd, __pmd(__pa(ptepage) | pmd_flags)); + } +#endif + + /* Get the pointer to the shadow PTE entry we're going to set. */ + return spte_addr(cpu, *spgd, vaddr); } /*H:330 @@ -204,60 +373,98 @@ static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd) * and return to the Guest without it knowing. * * If we fixed up the fault (ie. we mapped the address), this routine returns - * true. Otherwise, it was a real fault and we need to tell the Guest. */ -int demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) + * true. Otherwise, it was a real fault and we need to tell the Guest. + */ +bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) { - pgd_t gpgd; - pgd_t *spgd; unsigned long gpte_ptr; pte_t gpte; pte_t *spte; + pmd_t gpmd; + pgd_t gpgd; + + /* We never demand page the Switcher, so trying is a mistake. */ + if (vaddr >= switcher_addr) + return false; /* First step: get the top-level Guest page table entry. */ - gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); - /* Toplevel not present? We can't map it in. */ - if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) - return 0; + if (unlikely(cpu->linear_pages)) { + /* Faking up a linear mapping. */ + gpgd = __pgd(CHECK_GPGD_MASK); + } else { + gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); + /* Toplevel not present? We can't map it in. */ + if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) + return false; + + /* + * This kills the Guest if it has weird flags or tries to + * refer to a "physical" address outside the bounds. + */ + if (!check_gpgd(cpu, gpgd)) + return false; + } - /* Now look at the matching shadow entry. */ - spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); - if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) { - /* No shadow entry: allocate a new shadow PTE page. */ - unsigned long ptepage = get_zeroed_page(GFP_KERNEL); - /* This is not really the Guest's fault, but killing it is - * simple for this corner case. */ - if (!ptepage) { - kill_guest(cpu, "out of memory allocating pte page"); - return 0; - } - /* We check that the Guest pgd is OK. */ - check_gpgd(cpu, gpgd); - /* And we copy the flags to the shadow PGD entry. The page - * number in the shadow PGD is the page we just allocated. */ - *spgd = __pgd(__pa(ptepage) | pgd_flags(gpgd)); + /* This "mid-level" entry is only used for non-linear, PAE mode. */ + gpmd = __pmd(_PAGE_TABLE); + +#ifdef CONFIG_X86_PAE + if (likely(!cpu->linear_pages)) { + gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); + /* Middle level not present? We can't map it in. */ + if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) + return false; + + /* + * This kills the Guest if it has weird flags or tries to + * refer to a "physical" address outside the bounds. + */ + if (!check_gpmd(cpu, gpmd)) + return false; } - /* OK, now we look at the lower level in the Guest page table: keep its - * address, because we might update it later. */ - gpte_ptr = gpte_addr(gpgd, vaddr); - gpte = lgread(cpu, gpte_ptr, pte_t); + /* + * OK, now we look at the lower level in the Guest page table: keep its + * address, because we might update it later. + */ + gpte_ptr = gpte_addr(cpu, gpmd, vaddr); +#else + /* + * OK, now we look at the lower level in the Guest page table: keep its + * address, because we might update it later. + */ + gpte_ptr = gpte_addr(cpu, gpgd, vaddr); +#endif + + if (unlikely(cpu->linear_pages)) { + /* Linear? Make up a PTE which points to same page. */ + gpte = __pte((vaddr & PAGE_MASK) | _PAGE_RW | _PAGE_PRESENT); + } else { + /* Read the actual PTE value. */ + gpte = lgread(cpu, gpte_ptr, pte_t); + } /* If this page isn't in the Guest page tables, we can't page it in. */ if (!(pte_flags(gpte) & _PAGE_PRESENT)) - return 0; + return false; - /* Check they're not trying to write to a page the Guest wants - * read-only (bit 2 of errcode == write). */ + /* + * Check they're not trying to write to a page the Guest wants + * read-only (bit 2 of errcode == write). + */ if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW)) - return 0; + return false; /* User access to a kernel-only page? (bit 3 == user access) */ if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER)) - return 0; + return false; - /* Check that the Guest PTE flags are OK, and the page number is below - * the pfn_limit (ie. not mapping the Launcher binary). */ - check_gpte(cpu, gpte); + /* + * Check that the Guest PTE flags are OK, and the page number is below + * the pfn_limit (ie. not mapping the Launcher binary). + */ + if (!check_gpte(cpu, gpte)) + return false; /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */ gpte = pte_mkyoung(gpte); @@ -265,31 +472,45 @@ int demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) gpte = pte_mkdirty(gpte); /* Get the pointer to the shadow PTE entry we're going to set. */ - spte = spte_addr(*spgd, vaddr); - /* If there was a valid shadow PTE entry here before, we release it. - * This can happen with a write to a previously read-only entry. */ + spte = find_spte(cpu, vaddr, true, pgd_flags(gpgd), pmd_flags(gpmd)); + if (!spte) + return false; + + /* + * If there was a valid shadow PTE entry here before, we release it. + * This can happen with a write to a previously read-only entry. + */ release_pte(*spte); - /* If this is a write, we insist that the Guest page is writable (the - * final arg to gpte_to_spte()). */ + /* + * If this is a write, we insist that the Guest page is writable (the + * final arg to gpte_to_spte()). + */ if (pte_dirty(gpte)) *spte = gpte_to_spte(cpu, gpte, 1); else - /* If this is a read, don't set the "writable" bit in the page + /* + * If this is a read, don't set the "writable" bit in the page * table entry, even if the Guest says it's writable. That way * we will come back here when a write does actually occur, so - * we can update the Guest's _PAGE_DIRTY flag. */ - *spte = gpte_to_spte(cpu, pte_wrprotect(gpte), 0); - - /* Finally, we write the Guest PTE entry back: we've set the - * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */ - lgwrite(cpu, gpte_ptr, pte_t, gpte); - - /* The fault is fixed, the page table is populated, the mapping + * we can update the Guest's _PAGE_DIRTY flag. + */ + set_pte(spte, gpte_to_spte(cpu, pte_wrprotect(gpte), 0)); + + /* + * Finally, we write the Guest PTE entry back: we've set the + * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. + */ + if (likely(!cpu->linear_pages)) + lgwrite(cpu, gpte_ptr, pte_t, gpte); + + /* + * The fault is fixed, the page table is populated, the mapping * manipulated, the result returned and the code complete. A small * delay and a trace of alliteration are the only indications the Guest - * has that a page fault occurred at all. */ - return 1; + * has that a page fault occurred at all. + */ + return true; } /*H:360 @@ -301,42 +522,92 @@ int demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) * mapped, so it's overkill. * * This is a quick version which answers the question: is this virtual address - * mapped by the shadow page tables, and is it writable? */ -static int page_writable(struct lg_cpu *cpu, unsigned long vaddr) + * mapped by the shadow page tables, and is it writable? + */ +static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr) { - pgd_t *spgd; + pte_t *spte; unsigned long flags; - /* Look at the current top level entry: is it present? */ - spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); - if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) - return 0; + /* You can't put your stack in the Switcher! */ + if (vaddr >= switcher_addr) + return false; - /* Check the flags on the pte entry itself: it must be present and - * writable. */ - flags = pte_flags(*(spte_addr(*spgd, vaddr))); + /* If there's no shadow PTE, it's not writable. */ + spte = find_spte(cpu, vaddr, false, 0, 0); + if (!spte) + return false; + /* + * Check the flags on the pte entry itself: it must be present and + * writable. + */ + flags = pte_flags(*spte); return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); } -/* So, when pin_stack_pages() asks us to pin a page, we check if it's already +/* + * So, when pin_stack_pages() asks us to pin a page, we check if it's already * in the page tables, and if not, we call demand_page() with error code 2 - * (meaning "write"). */ + * (meaning "write"). + */ void pin_page(struct lg_cpu *cpu, unsigned long vaddr) { if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2)) kill_guest(cpu, "bad stack page %#lx", vaddr); } +/*:*/ + +#ifdef CONFIG_X86_PAE +static void release_pmd(pmd_t *spmd) +{ + /* If the entry's not present, there's nothing to release. */ + if (pmd_flags(*spmd) & _PAGE_PRESENT) { + unsigned int i; + pte_t *ptepage = __va(pmd_pfn(*spmd) << PAGE_SHIFT); + /* For each entry in the page, we might need to release it. */ + for (i = 0; i < PTRS_PER_PTE; i++) + release_pte(ptepage[i]); + /* Now we can free the page of PTEs */ + free_page((long)ptepage); + /* And zero out the PMD entry so we never release it twice. */ + set_pmd(spmd, __pmd(0)); + } +} + +static void release_pgd(pgd_t *spgd) +{ + /* If the entry's not present, there's nothing to release. */ + if (pgd_flags(*spgd) & _PAGE_PRESENT) { + unsigned int i; + pmd_t *pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); + + for (i = 0; i < PTRS_PER_PMD; i++) + release_pmd(&pmdpage[i]); + + /* Now we can free the page of PMDs */ + free_page((long)pmdpage); + /* And zero out the PGD entry so we never release it twice. */ + set_pgd(spgd, __pgd(0)); + } +} -/*H:450 If we chase down the release_pgd() code, it looks like this: */ -static void release_pgd(struct lguest *lg, pgd_t *spgd) +#else /* !CONFIG_X86_PAE */ +/*H:450 + * If we chase down the release_pgd() code, the non-PAE version looks like + * this. The PAE version is almost identical, but instead of calling + * release_pte it calls release_pmd(), which looks much like this. + */ +static void release_pgd(pgd_t *spgd) { /* If the entry's not present, there's nothing to release. */ if (pgd_flags(*spgd) & _PAGE_PRESENT) { unsigned int i; - /* Converting the pfn to find the actual PTE page is easy: turn + /* + * Converting the pfn to find the actual PTE page is easy: turn * the page number into a physical address, then convert to a - * virtual address (easy for kernel pages like this one). */ + * virtual address (easy for kernel pages like this one). + */ pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); /* For each entry in the page, we might need to release it. */ for (i = 0; i < PTRS_PER_PTE; i++) @@ -347,22 +618,27 @@ static void release_pgd(struct lguest *lg, pgd_t *spgd) *spgd = __pgd(0); } } +#endif -/*H:445 We saw flush_user_mappings() twice: once from the flush_user_mappings() +/*H:445 + * We saw flush_user_mappings() twice: once from the flush_user_mappings() * hypercall and once in new_pgdir() when we re-used a top-level pgdir page. - * It simply releases every PTE page from 0 up to the Guest's kernel address. */ + * It simply releases every PTE page from 0 up to the Guest's kernel address. + */ static void flush_user_mappings(struct lguest *lg, int idx) { unsigned int i; /* Release every pgd entry up to the kernel's address. */ for (i = 0; i < pgd_index(lg->kernel_address); i++) - release_pgd(lg, lg->pgdirs[idx].pgdir + i); + release_pgd(lg->pgdirs[idx].pgdir + i); } -/*H:440 (v) Flushing (throwing away) page tables, +/*H:440 + * (v) Flushing (throwing away) page tables, * * The Guest has a hypercall to throw away the page tables: it's used when a - * large number of mappings have been changed. */ + * large number of mappings have been changed. + */ void guest_pagetable_flush_user(struct lg_cpu *cpu) { /* Drop the userspace part of the current page table. */ @@ -375,23 +651,43 @@ unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr) { pgd_t gpgd; pte_t gpte; +#ifdef CONFIG_X86_PAE + pmd_t gpmd; +#endif + + /* Still not set up? Just map 1:1. */ + if (unlikely(cpu->linear_pages)) + return vaddr; /* First step: get the top-level Guest page table entry. */ gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); /* Toplevel not present? We can't map it in. */ - if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) + if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) { kill_guest(cpu, "Bad address %#lx", vaddr); + return -1UL; + } - gpte = lgread(cpu, gpte_addr(gpgd, vaddr), pte_t); +#ifdef CONFIG_X86_PAE + gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); + if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) { + kill_guest(cpu, "Bad address %#lx", vaddr); + return -1UL; + } + gpte = lgread(cpu, gpte_addr(cpu, gpmd, vaddr), pte_t); +#else + gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t); +#endif if (!(pte_flags(gpte) & _PAGE_PRESENT)) kill_guest(cpu, "Bad address %#lx", vaddr); return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK); } -/* We keep several page tables. This is a simple routine to find the page +/* + * We keep several page tables. This is a simple routine to find the page * table (if any) corresponding to this top-level address the Guest has given - * us. */ + * us. + */ static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) { unsigned int i; @@ -401,18 +697,22 @@ static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) return i; } -/*H:435 And this is us, creating the new page directory. If we really do +/*H:435 + * And this is us, creating the new page directory. If we really do * allocate a new one (and so the kernel parts are not there), we set - * blank_pgdir. */ + * blank_pgdir. + */ static unsigned int new_pgdir(struct lg_cpu *cpu, unsigned long gpgdir, int *blank_pgdir) { unsigned int next; - /* We pick one entry at random to throw out. Choosing the Least - * Recently Used might be better, but this is easy. */ - next = random32() % ARRAY_SIZE(cpu->lg->pgdirs); + /* + * We pick one entry at random to throw out. Choosing the Least + * Recently Used might be better, but this is easy. + */ + next = prandom_u32() % ARRAY_SIZE(cpu->lg->pgdirs); /* If it's never been allocated at all before, try now. */ if (!cpu->lg->pgdirs[next].pgdir) { cpu->lg->pgdirs[next].pgdir = @@ -420,76 +720,160 @@ static unsigned int new_pgdir(struct lg_cpu *cpu, /* If the allocation fails, just keep using the one we have */ if (!cpu->lg->pgdirs[next].pgdir) next = cpu->cpu_pgd; - else - /* This is a blank page, so there are no kernel - * mappings: caller must map the stack! */ + else { + /* + * This is a blank page, so there are no kernel + * mappings: caller must map the stack! + */ *blank_pgdir = 1; + } } /* Record which Guest toplevel this shadows. */ cpu->lg->pgdirs[next].gpgdir = gpgdir; /* Release all the non-kernel mappings. */ flush_user_mappings(cpu->lg, next); + /* This hasn't run on any CPU at all. */ + cpu->lg->pgdirs[next].last_host_cpu = -1; + return next; } -/*H:430 (iv) Switching page tables +/*H:501 + * We do need the Switcher code mapped at all times, so we allocate that + * part of the Guest page table here. We map the Switcher code immediately, + * but defer mapping of the guest register page and IDT/LDT etc page until + * just before we run the guest in map_switcher_in_guest(). * - * Now we've seen all the page table setting and manipulation, let's see what - * what happens when the Guest changes page tables (ie. changes the top-level - * pgdir). This occurs on almost every context switch. */ -void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable) + * We *could* do this setup in map_switcher_in_guest(), but at that point + * we've interrupts disabled, and allocating pages like that is fraught: we + * can't sleep if we need to free up some memory. + */ +static bool allocate_switcher_mapping(struct lg_cpu *cpu) { - int newpgdir, repin = 0; - - /* Look to see if we have this one already. */ - newpgdir = find_pgdir(cpu->lg, pgtable); - /* If not, we allocate or mug an existing one: if it's a fresh one, - * repin gets set to 1. */ - if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs)) - newpgdir = new_pgdir(cpu, pgtable, &repin); - /* Change the current pgd index to the new one. */ - cpu->cpu_pgd = newpgdir; - /* If it was completely blank, we map in the Guest kernel stack */ - if (repin) - pin_stack_pages(cpu); + int i; + + for (i = 0; i < TOTAL_SWITCHER_PAGES; i++) { + pte_t *pte = find_spte(cpu, switcher_addr + i * PAGE_SIZE, true, + CHECK_GPGD_MASK, _PAGE_TABLE); + if (!pte) + return false; + + /* + * Map the switcher page if not already there. It might + * already be there because we call allocate_switcher_mapping() + * in guest_set_pgd() just in case it did discard our Switcher + * mapping, but it probably didn't. + */ + if (i == 0 && !(pte_flags(*pte) & _PAGE_PRESENT)) { + /* Get a reference to the Switcher page. */ + get_page(lg_switcher_pages[0]); + /* Create a read-only, exectuable, kernel-style PTE */ + set_pte(pte, + mk_pte(lg_switcher_pages[0], PAGE_KERNEL_RX)); + } + } + cpu->lg->pgdirs[cpu->cpu_pgd].switcher_mapped = true; + return true; } -/*H:470 Finally, a routine which throws away everything: all PGD entries in all +/*H:470 + * Finally, a routine which throws away everything: all PGD entries in all * the shadow page tables, including the Guest's kernel mappings. This is used - * when we destroy the Guest. */ + * when we destroy the Guest. + */ static void release_all_pagetables(struct lguest *lg) { unsigned int i, j; /* Every shadow pagetable this Guest has */ - for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) - if (lg->pgdirs[i].pgdir) - /* Every PGD entry except the Switcher at the top */ - for (j = 0; j < SWITCHER_PGD_INDEX; j++) - release_pgd(lg, lg->pgdirs[i].pgdir + j); + for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) { + if (!lg->pgdirs[i].pgdir) + continue; + + /* Every PGD entry. */ + for (j = 0; j < PTRS_PER_PGD; j++) + release_pgd(lg->pgdirs[i].pgdir + j); + lg->pgdirs[i].switcher_mapped = false; + lg->pgdirs[i].last_host_cpu = -1; + } } -/* We also throw away everything when a Guest tells us it's changed a kernel +/* + * We also throw away everything when a Guest tells us it's changed a kernel * mapping. Since kernel mappings are in every page table, it's easiest to * throw them all away. This traps the Guest in amber for a while as - * everything faults back in, but it's rare. */ + * everything faults back in, but it's rare. + */ void guest_pagetable_clear_all(struct lg_cpu *cpu) { release_all_pagetables(cpu->lg); /* We need the Guest kernel stack mapped again. */ pin_stack_pages(cpu); + /* And we need Switcher allocated. */ + if (!allocate_switcher_mapping(cpu)) + kill_guest(cpu, "Cannot populate switcher mapping"); +} + +/*H:430 + * (iv) Switching page tables + * + * Now we've seen all the page table setting and manipulation, let's see + * what happens when the Guest changes page tables (ie. changes the top-level + * pgdir). This occurs on almost every context switch. + */ +void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable) +{ + int newpgdir, repin = 0; + + /* + * The very first time they call this, we're actually running without + * any page tables; we've been making it up. Throw them away now. + */ + if (unlikely(cpu->linear_pages)) { + release_all_pagetables(cpu->lg); + cpu->linear_pages = false; + /* Force allocation of a new pgdir. */ + newpgdir = ARRAY_SIZE(cpu->lg->pgdirs); + } else { + /* Look to see if we have this one already. */ + newpgdir = find_pgdir(cpu->lg, pgtable); + } + + /* + * If not, we allocate or mug an existing one: if it's a fresh one, + * repin gets set to 1. + */ + if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs)) + newpgdir = new_pgdir(cpu, pgtable, &repin); + /* Change the current pgd index to the new one. */ + cpu->cpu_pgd = newpgdir; + /* + * If it was completely blank, we map in the Guest kernel stack and + * the Switcher. + */ + if (repin) + pin_stack_pages(cpu); + + if (!cpu->lg->pgdirs[cpu->cpu_pgd].switcher_mapped) { + if (!allocate_switcher_mapping(cpu)) + kill_guest(cpu, "Cannot populate switcher mapping"); + } } /*:*/ -/*M:009 Since we throw away all mappings when a kernel mapping changes, our + +/*M:009 + * Since we throw away all mappings when a kernel mapping changes, our * performance sucks for guests using highmem. In fact, a guest with * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is * usually slower than a Guest with less memory. * * This, of course, cannot be fixed. It would take some kind of... well, I - * don't know, but the term "puissant code-fu" comes to mind. :*/ + * don't know, but the term "puissant code-fu" comes to mind. +:*/ -/*H:420 This is the routine which actually sets the page table entry for then +/*H:420 + * This is the routine which actually sets the page table entry for then * "idx"'th shadow page table. * * Normally, we can just throw out the old entry and replace it with 0: if they @@ -503,33 +887,52 @@ void guest_pagetable_clear_all(struct lg_cpu *cpu) * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately. */ -static void do_set_pte(struct lg_cpu *cpu, int idx, +static void __guest_set_pte(struct lg_cpu *cpu, int idx, unsigned long vaddr, pte_t gpte) { /* Look up the matching shadow page directory entry. */ pgd_t *spgd = spgd_addr(cpu, idx, vaddr); +#ifdef CONFIG_X86_PAE + pmd_t *spmd; +#endif /* If the top level isn't present, there's no entry to update. */ if (pgd_flags(*spgd) & _PAGE_PRESENT) { - /* Otherwise, we start by releasing the existing entry. */ - pte_t *spte = spte_addr(*spgd, vaddr); - release_pte(*spte); - - /* If they're setting this entry as dirty or accessed, we might - * as well put that entry they've given us in now. This shaves - * 10% off a copy-on-write micro-benchmark. */ - if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) { - check_gpte(cpu, gpte); - *spte = gpte_to_spte(cpu, gpte, - pte_flags(gpte) & _PAGE_DIRTY); - } else - /* Otherwise kill it and we can demand_page() it in - * later. */ - *spte = __pte(0); +#ifdef CONFIG_X86_PAE + spmd = spmd_addr(cpu, *spgd, vaddr); + if (pmd_flags(*spmd) & _PAGE_PRESENT) { +#endif + /* Otherwise, start by releasing the existing entry. */ + pte_t *spte = spte_addr(cpu, *spgd, vaddr); + release_pte(*spte); + + /* + * If they're setting this entry as dirty or accessed, + * we might as well put that entry they've given us in + * now. This shaves 10% off a copy-on-write + * micro-benchmark. + */ + if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) { + if (!check_gpte(cpu, gpte)) + return; + set_pte(spte, + gpte_to_spte(cpu, gpte, + pte_flags(gpte) & _PAGE_DIRTY)); + } else { + /* + * Otherwise kill it and we can demand_page() + * it in later. + */ + set_pte(spte, __pte(0)); + } +#ifdef CONFIG_X86_PAE + } +#endif } } -/*H:410 Updating a PTE entry is a little trickier. +/*H:410 + * Updating a PTE entry is a little trickier. * * We keep track of several different page tables (the Guest uses one for each * process, so it makes sense to cache at least a few). Each of these have @@ -538,23 +941,32 @@ static void do_set_pte(struct lg_cpu *cpu, int idx, * all the page tables, not just the current one. This is rare. * * The benefit is that when we have to track a new page table, we can keep all - * the kernel mappings. This speeds up context switch immensely. */ + * the kernel mappings. This speeds up context switch immensely. + */ void guest_set_pte(struct lg_cpu *cpu, unsigned long gpgdir, unsigned long vaddr, pte_t gpte) { - /* Kernel mappings must be changed on all top levels. Slow, but doesn't - * happen often. */ + /* We don't let you remap the Switcher; we need it to get back! */ + if (vaddr >= switcher_addr) { + kill_guest(cpu, "attempt to set pte into Switcher pages"); + return; + } + + /* + * Kernel mappings must be changed on all top levels. Slow, but doesn't + * happen often. + */ if (vaddr >= cpu->lg->kernel_address) { unsigned int i; for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++) if (cpu->lg->pgdirs[i].pgdir) - do_set_pte(cpu, i, vaddr, gpte); + __guest_set_pte(cpu, i, vaddr, gpte); } else { /* Is this page table one we have a shadow for? */ int pgdir = find_pgdir(cpu->lg, gpgdir); if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs)) /* If so, do the update. */ - do_set_pte(cpu, pgdir, vaddr, gpte); + __guest_set_pte(cpu, pgdir, vaddr, gpte); } } @@ -570,56 +982,101 @@ void guest_set_pte(struct lg_cpu *cpu, * tells us they've changed. When the Guest tries to use the new entry it will * fault and demand_page() will fix it up. * - * So with that in mind here's our code to to update a (top-level) PGD entry: + * So with that in mind here's our code to update a (top-level) PGD entry: */ -void guest_set_pmd(struct lguest *lg, unsigned long gpgdir, u32 idx) +void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx) { int pgdir; - /* The kernel seems to try to initialize this early on: we ignore its - * attempts to map over the Switcher. */ - if (idx >= SWITCHER_PGD_INDEX) + if (idx > PTRS_PER_PGD) { + kill_guest(&lg->cpus[0], "Attempt to set pgd %u/%u", + idx, PTRS_PER_PGD); return; + } /* If they're talking about a page table we have a shadow for... */ pgdir = find_pgdir(lg, gpgdir); - if (pgdir < ARRAY_SIZE(lg->pgdirs)) + if (pgdir < ARRAY_SIZE(lg->pgdirs)) { /* ... throw it away. */ - release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx); + release_pgd(lg->pgdirs[pgdir].pgdir + idx); + /* That might have been the Switcher mapping, remap it. */ + if (!allocate_switcher_mapping(&lg->cpus[0])) { + kill_guest(&lg->cpus[0], + "Cannot populate switcher mapping"); + } + lg->pgdirs[pgdir].last_host_cpu = -1; + } } -/*H:500 (vii) Setting up the page tables initially. +#ifdef CONFIG_X86_PAE +/* For setting a mid-level, we just throw everything away. It's easy. */ +void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx) +{ + guest_pagetable_clear_all(&lg->cpus[0]); +} +#endif + +/*H:500 + * (vii) Setting up the page tables initially. + * + * When a Guest is first created, set initialize a shadow page table which + * we will populate on future faults. The Guest doesn't have any actual + * pagetables yet, so we set linear_pages to tell demand_page() to fake it + * for the moment. * - * When a Guest is first created, the Launcher tells us where the toplevel of - * its first page table is. We set some things up here: */ -int init_guest_pagetable(struct lguest *lg, unsigned long pgtable) -{ - /* We start on the first shadow page table, and give it a blank PGD - * page. */ - lg->pgdirs[0].gpgdir = pgtable; - lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL); - if (!lg->pgdirs[0].pgdir) + * We do need the Switcher to be mapped at all times, so we allocate that + * part of the Guest page table here. + */ +int init_guest_pagetable(struct lguest *lg) +{ + struct lg_cpu *cpu = &lg->cpus[0]; + int allocated = 0; + + /* lg (and lg->cpus[]) starts zeroed: this allocates a new pgdir */ + cpu->cpu_pgd = new_pgdir(cpu, 0, &allocated); + if (!allocated) return -ENOMEM; - lg->cpus[0].cpu_pgd = 0; + + /* We start with a linear mapping until the initialize. */ + cpu->linear_pages = true; + + /* Allocate the page tables for the Switcher. */ + if (!allocate_switcher_mapping(cpu)) { + release_all_pagetables(lg); + return -ENOMEM; + } + return 0; } -/* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ +/*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ void page_table_guest_data_init(struct lg_cpu *cpu) { + /* + * We tell the Guest that it can't use the virtual addresses + * used by the Switcher. This trick is equivalent to 4GB - + * switcher_addr. + */ + u32 top = ~switcher_addr + 1; + /* We get the kernel address: above this is all kernel memory. */ if (get_user(cpu->lg->kernel_address, &cpu->lg->lguest_data->kernel_address) - /* We tell the Guest that it can't use the top 4MB of virtual - * addresses used by the Switcher. */ - || put_user(4U*1024*1024, &cpu->lg->lguest_data->reserve_mem) - || put_user(cpu->lg->pgdirs[0].gpgdir, &cpu->lg->lguest_data->pgdir)) + /* + * We tell the Guest that it can't use the top virtual + * addresses (used by the Switcher). + */ + || put_user(top, &cpu->lg->lguest_data->reserve_mem)) { kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data); + return; + } - /* In flush_user_mappings() we loop from 0 to + /* + * In flush_user_mappings() we loop from 0 to * "pgd_index(lg->kernel_address)". This assumes it won't hit the - * Switcher mappings, so check that now. */ - if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX) + * Switcher mappings, so check that now. + */ + if (cpu->lg->kernel_address >= switcher_addr) kill_guest(cpu, "bad kernel address %#lx", cpu->lg->kernel_address); } @@ -636,77 +1093,97 @@ void free_guest_pagetable(struct lguest *lg) free_page((long)lg->pgdirs[i].pgdir); } -/*H:480 (vi) Mapping the Switcher when the Guest is about to run. - * - * The Switcher and the two pages for this CPU need to be visible in the - * Guest (and not the pages for other CPUs). We have the appropriate PTE pages - * for each CPU already set up, we just need to hook them in now we know which - * Guest is about to run on this CPU. */ -void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages) +/*H:481 + * This clears the Switcher mappings for cpu #i. + */ +static void remove_switcher_percpu_map(struct lg_cpu *cpu, unsigned int i) { - pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages); - pgd_t switcher_pgd; - pte_t regs_pte; - unsigned long pfn; - - /* Make the last PGD entry for this Guest point to the Switcher's PTE - * page for this CPU (with appropriate flags). */ - switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL); - - cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; + unsigned long base = switcher_addr + PAGE_SIZE + i * PAGE_SIZE*2; + pte_t *pte; - /* We also change the Switcher PTE page. When we're running the Guest, - * we want the Guest's "regs" page to appear where the first Switcher - * page for this CPU is. This is an optimization: when the Switcher - * saves the Guest registers, it saves them into the first page of this - * CPU's "struct lguest_pages": if we make sure the Guest's register - * page is already mapped there, we don't have to copy them out - * again. */ - pfn = __pa(cpu->regs_page) >> PAGE_SHIFT; - regs_pte = pfn_pte(pfn, __pgprot(__PAGE_KERNEL)); - switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTRS_PER_PTE] = regs_pte; -} -/*:*/ + /* Clear the mappings for both pages. */ + pte = find_spte(cpu, base, false, 0, 0); + release_pte(*pte); + set_pte(pte, __pte(0)); -static void free_switcher_pte_pages(void) -{ - unsigned int i; - - for_each_possible_cpu(i) - free_page((long)switcher_pte_page(i)); + pte = find_spte(cpu, base + PAGE_SIZE, false, 0, 0); + release_pte(*pte); + set_pte(pte, __pte(0)); } -/*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given - * the CPU number and the "struct page"s for the Switcher code itself. +/*H:480 + * (vi) Mapping the Switcher when the Guest is about to run. + * + * The Switcher and the two pages for this CPU need to be visible in the Guest + * (and not the pages for other CPUs). * - * Currently the Switcher is less than a page long, so "pages" is always 1. */ -static __init void populate_switcher_pte_page(unsigned int cpu, - struct page *switcher_page[], - unsigned int pages) + * The pages for the pagetables have all been allocated before: we just need + * to make sure the actual PTEs are up-to-date for the CPU we're about to run + * on. + */ +void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages) { - unsigned int i; - pte_t *pte = switcher_pte_page(cpu); + unsigned long base; + struct page *percpu_switcher_page, *regs_page; + pte_t *pte; + struct pgdir *pgdir = &cpu->lg->pgdirs[cpu->cpu_pgd]; + + /* Switcher page should always be mapped by now! */ + BUG_ON(!pgdir->switcher_mapped); + + /* + * Remember that we have two pages for each Host CPU, so we can run a + * Guest on each CPU without them interfering. We need to make sure + * those pages are mapped correctly in the Guest, but since we usually + * run on the same CPU, we cache that, and only update the mappings + * when we move. + */ + if (pgdir->last_host_cpu == raw_smp_processor_id()) + return; - /* The first entries are easy: they map the Switcher code. */ - for (i = 0; i < pages; i++) { - pte[i] = mk_pte(switcher_page[i], - __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)); + /* -1 means unknown so we remove everything. */ + if (pgdir->last_host_cpu == -1) { + unsigned int i; + for_each_possible_cpu(i) + remove_switcher_percpu_map(cpu, i); + } else { + /* We know exactly what CPU mapping to remove. */ + remove_switcher_percpu_map(cpu, pgdir->last_host_cpu); } - /* The only other thing we map is this CPU's pair of pages. */ - i = pages + cpu*2; - - /* First page (Guest registers) is writable from the Guest */ - pte[i] = pfn_pte(page_to_pfn(switcher_page[i]), - __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW)); - - /* The second page contains the "struct lguest_ro_state", and is - * read-only. */ - pte[i+1] = pfn_pte(page_to_pfn(switcher_page[i+1]), - __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)); + /* + * When we're running the Guest, we want the Guest's "regs" page to + * appear where the first Switcher page for this CPU is. This is an + * optimization: when the Switcher saves the Guest registers, it saves + * them into the first page of this CPU's "struct lguest_pages": if we + * make sure the Guest's register page is already mapped there, we + * don't have to copy them out again. + */ + /* Find the shadow PTE for this regs page. */ + base = switcher_addr + PAGE_SIZE + + raw_smp_processor_id() * sizeof(struct lguest_pages); + pte = find_spte(cpu, base, false, 0, 0); + regs_page = pfn_to_page(__pa(cpu->regs_page) >> PAGE_SHIFT); + get_page(regs_page); + set_pte(pte, mk_pte(regs_page, __pgprot(__PAGE_KERNEL & ~_PAGE_GLOBAL))); + + /* + * We map the second page of the struct lguest_pages read-only in + * the Guest: the IDT, GDT and other things it's not supposed to + * change. + */ + pte = find_spte(cpu, base + PAGE_SIZE, false, 0, 0); + percpu_switcher_page + = lg_switcher_pages[1 + raw_smp_processor_id()*2 + 1]; + get_page(percpu_switcher_page); + set_pte(pte, mk_pte(percpu_switcher_page, + __pgprot(__PAGE_KERNEL_RO & ~_PAGE_GLOBAL))); + + pgdir->last_host_cpu = raw_smp_processor_id(); } -/* We've made it through the page table code. Perhaps our tired brains are +/*H:490 + * We've made it through the page table code. Perhaps our tired brains are * still processing the details, or perhaps we're simply glad it's over. * * If nothing else, note that all this complexity in juggling shadow page tables @@ -715,28 +1192,5 @@ static __init void populate_switcher_pte_page(unsigned int cpu, * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD * have implemented shadow page table support directly into hardware. * - * There is just one file remaining in the Host. */ - -/*H:510 At boot or module load time, init_pagetables() allocates and populates - * the Switcher PTE page for each CPU. */ -__init int init_pagetables(struct page **switcher_page, unsigned int pages) -{ - unsigned int i; - - for_each_possible_cpu(i) { - switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL); - if (!switcher_pte_page(i)) { - free_switcher_pte_pages(); - return -ENOMEM; - } - populate_switcher_pte_page(i, switcher_page, pages); - } - return 0; -} -/*:*/ - -/* Cleaning up simply involves freeing the PTE page for each CPU. */ -void free_pagetables(void) -{ - free_switcher_pte_pages(); -} + * There is just one file remaining in the Host. + */ |