X-Git-Url: https://err.no/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=drivers%2Flguest%2Fpage_tables.c;h=fffabb3271573d4f979cb36e98b5103c1d1246a7;hb=e48d33193d94175f012c3ed606a1d1e574ed726a;hp=b7a924ace68426d85167adb6121f7e39364e9be4;hpb=644b55ce889edd37d6406df26e2d96d7a7390749;p=linux-2.6 diff --git a/drivers/lguest/page_tables.c b/drivers/lguest/page_tables.c index b7a924ace6..fffabb3271 100644 --- a/drivers/lguest/page_tables.c +++ b/drivers/lguest/page_tables.c @@ -13,6 +13,7 @@ #include #include #include +#include #include "lg.h" /*M:008 We hold reference to pages, which prevents them from being swapped. @@ -25,7 +26,8 @@ * * 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 lguest.c's "Page Table Handling" (with diagrams!). + * 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 @@ -35,53 +37,40 @@ * * Anyway, this is the most complicated part of the Host code. There are seven * parts to this: - * (i) Setting up a page table entry for the Guest when it faults, - * (ii) Setting up the page table entry for the Guest stack, - * (iii) Setting up a page table entry when the Guest tells us it has changed, + * (i) Looking up a page table entry when the Guest faults, + * (ii) Making sure the Guest stack is mapped, + * (iii) Setting up a page table entry when the Guest tells us one has changed, * (iv) Switching page tables, - * (v) Flushing (thowing away) page tables, + * (v) Flushing (throwing away) page tables, * (vi) Mapping the Switcher when the Guest is about to run, * (vii) Setting up the page tables initially. :*/ -/* Pages a 4k long, and each page table entry is 4 bytes long, giving us 1024 - * (or 2^10) entries per page. */ -#define PTES_PER_PAGE_SHIFT 10 -#define PTES_PER_PAGE (1 << PTES_PER_PAGE_SHIFT) /* 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. */ -#define SWITCHER_PGD_INDEX (PTES_PER_PAGE - 1) +#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(spte_t *, switcher_pte_pages); +static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) -/*H:320 With our shadow and Guest types established, we need to deal with - * them: the page table code is curly enough to need helper functions to keep - * it clear and clean. +/*H:320 The page table code is curly enough to need helper functions to keep it + * clear and clean. * - * The first helper takes a virtual address, and says which entry in the top - * level page table deals with that address. Since each top level entry deals - * with 4M, this effectively divides by 4M. */ -static unsigned vaddr_to_pgd_index(unsigned long vaddr) -{ - return vaddr >> (PAGE_SHIFT + PTES_PER_PAGE_SHIFT); -} - -/* There are two functions which return pointers to the shadow (aka "real") + * There are two functions which return pointers to the shadow (aka "real") * page tables. * * spgd_addr() takes the virtual address and returns a pointer to the top-level - * page directory entry for that address. Since we keep track of several page - * tables, the "i" argument tells us which one we're interested in (it's + * 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). */ -static spgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr) +static pgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr) { - unsigned int index = vaddr_to_pgd_index(vaddr); + unsigned int index = pgd_index(vaddr); /* We kill any Guest trying to touch the Switcher addresses. */ if (index >= SWITCHER_PGD_INDEX) { @@ -92,31 +81,31 @@ static spgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr) return &lg->pgdirs[i].pgdir[index]; } -/* This routine then takes the PGD entry given above, which contains the - * address of the PTE page. It then returns a pointer to the PTE entry for the - * given address. */ -static spte_t *spte_addr(struct lguest *lg, spgd_t spgd, unsigned long vaddr) +/* 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(struct lguest *lg, pgd_t spgd, unsigned long vaddr) { - spte_t *page = __va(spgd.pfn << PAGE_SHIFT); + pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT); /* You should never call this if the PGD entry wasn't valid */ - BUG_ON(!(spgd.flags & _PAGE_PRESENT)); - return &page[(vaddr >> PAGE_SHIFT) % PTES_PER_PAGE]; + BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); + return &page[(vaddr >> PAGE_SHIFT) % PTRS_PER_PTE]; } /* These two functions just like the above two, except they access the Guest * page tables. Hence they return a Guest address. */ static unsigned long gpgd_addr(struct lguest *lg, unsigned long vaddr) { - unsigned int index = vaddr >> (PAGE_SHIFT + PTES_PER_PAGE_SHIFT); - return lg->pgdirs[lg->pgdidx].cr3 + index * sizeof(gpgd_t); + unsigned int index = vaddr >> (PGDIR_SHIFT); + return lg->pgdirs[lg->pgdidx].gpgdir + index * sizeof(pgd_t); } static unsigned long gpte_addr(struct lguest *lg, - gpgd_t gpgd, unsigned long vaddr) + pgd_t gpgd, unsigned long vaddr) { - unsigned long gpage = gpgd.pfn << PAGE_SHIFT; - BUG_ON(!(gpgd.flags & _PAGE_PRESENT)); - return gpage + ((vaddr>>PAGE_SHIFT) % PTES_PER_PAGE) * sizeof(gpte_t); + 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); } /*H:350 This routine takes a page number given by the Guest and converts it to @@ -149,58 +138,60 @@ static unsigned long get_pfn(unsigned long virtpfn, int write) * 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. */ -static spte_t gpte_to_spte(struct lguest *lg, gpte_t gpte, int write) +static pte_t gpte_to_spte(struct lguest *lg, pte_t gpte, int write) { - spte_t spte; - unsigned long pfn; + unsigned long pfn, base, flags; /* 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. */ - spte.flags = (gpte.flags & ~_PAGE_GLOBAL); + flags = (pte_flags(gpte) & ~_PAGE_GLOBAL); + + /* The Guest's pages are offset inside the Launcher. */ + base = (unsigned long)lg->mem_base / PAGE_SIZE; /* 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. */ - pfn = get_pfn(gpte.pfn, write); + pfn = get_pfn(base + pte_pfn(gpte), write); if (pfn == -1UL) { - kill_guest(lg, "failed to get page %u", gpte.pfn); + kill_guest(lg, "failed to get page %lu", pte_pfn(gpte)); /* 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! */ - spte.flags = 0; + flags = 0; } - /* Now we assign the page number, and our shadow PTE is complete. */ - spte.pfn = pfn; - return spte; + /* Now we assemble our shadow PTE from the page number and flags. */ + return pfn_pte(pfn, __pgprot(flags)); } /*H:460 And to complete the chain, release_pte() looks like this: */ -static void release_pte(spte_t pte) +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. */ - if (pte.flags & _PAGE_PRESENT) - put_page(pfn_to_page(pte.pfn)); + if (pte_flags(pte) & _PAGE_PRESENT) + put_page(pfn_to_page(pte_pfn(pte))); } /*:*/ -static void check_gpte(struct lguest *lg, gpte_t gpte) +static void check_gpte(struct lguest *lg, pte_t gpte) { - if ((gpte.flags & (_PAGE_PWT|_PAGE_PSE)) || gpte.pfn >= lg->pfn_limit) + if ((pte_flags(gpte) & (_PAGE_PWT|_PAGE_PSE)) + || pte_pfn(gpte) >= lg->pfn_limit) kill_guest(lg, "bad page table entry"); } -static void check_gpgd(struct lguest *lg, gpgd_t gpgd) +static void check_gpgd(struct lguest *lg, pgd_t gpgd) { - if ((gpgd.flags & ~_PAGE_TABLE) || gpgd.pfn >= lg->pfn_limit) + if ((pgd_flags(gpgd) & ~_PAGE_TABLE) || pgd_pfn(gpgd) >= lg->pfn_limit) kill_guest(lg, "bad page directory entry"); } /*H:330 - * (i) Setting up a page table entry for the Guest when it faults + * (i) Looking up a page table entry when the Guest faults. * * We saw this call in run_guest(): when we see a page fault in the Guest, we * come here. That's because we only set up the shadow page tables lazily as @@ -208,24 +199,24 @@ static void check_gpgd(struct lguest *lg, gpgd_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. */ + * true. Otherwise, it was a real fault and we need to tell the Guest. */ int demand_page(struct lguest *lg, unsigned long vaddr, int errcode) { - gpgd_t gpgd; - spgd_t *spgd; + pgd_t gpgd; + pgd_t *spgd; unsigned long gpte_ptr; - gpte_t gpte; - spte_t *spte; + pte_t gpte; + pte_t *spte; /* First step: get the top-level Guest page table entry. */ - gpgd = mkgpgd(lgread_u32(lg, gpgd_addr(lg, vaddr))); + gpgd = lgread(lg, gpgd_addr(lg, vaddr), pgd_t); /* Toplevel not present? We can't map it in. */ - if (!(gpgd.flags & _PAGE_PRESENT)) + if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) return 0; /* Now look at the matching shadow entry. */ spgd = spgd_addr(lg, lg->pgdidx, vaddr); - if (!(spgd->flags & _PAGE_PRESENT)) { + 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 @@ -238,34 +229,35 @@ int demand_page(struct lguest *lg, unsigned long vaddr, int errcode) check_gpgd(lg, 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->raw.val = (__pa(ptepage) | gpgd.flags); + *spgd = __pgd(__pa(ptepage) | pgd_flags(gpgd)); } /* 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(lg, gpgd, vaddr); - gpte = mkgpte(lgread_u32(lg, gpte_ptr)); + gpte = lgread(lg, gpte_ptr, pte_t); /* If this page isn't in the Guest page tables, we can't page it in. */ - if (!(gpte.flags & _PAGE_PRESENT)) + if (!(pte_flags(gpte) & _PAGE_PRESENT)) return 0; /* Check they're not trying to write to a page the Guest wants * read-only (bit 2 of errcode == write). */ - if ((errcode & 2) && !(gpte.flags & _PAGE_RW)) + if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW)) return 0; - /* User access to a kernel page? (bit 3 == user access) */ - if ((errcode & 4) && !(gpte.flags & _PAGE_USER)) + /* User access to a kernel-only page? (bit 3 == user access) */ + if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER)) return 0; /* 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(lg, gpte); + /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */ - gpte.flags |= _PAGE_ACCESSED; + gpte = pte_mkyoung(gpte); if (errcode & 2) - gpte.flags |= _PAGE_DIRTY; + gpte = pte_mkdirty(gpte); /* Get the pointer to the shadow PTE entry we're going to set. */ spte = spte_addr(lg, *spgd, vaddr); @@ -275,47 +267,50 @@ int demand_page(struct lguest *lg, unsigned long vaddr, int errcode) /* If this is a write, we insist that the Guest page is writable (the * final arg to gpte_to_spte()). */ - if (gpte.flags & _PAGE_DIRTY) + if (pte_dirty(gpte)) *spte = gpte_to_spte(lg, gpte, 1); - else { + else /* 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 come back here when a write does actually ocur, so we can - * update the Guest's _PAGE_DIRTY flag. */ - gpte_t ro_gpte = gpte; - ro_gpte.flags &= ~_PAGE_RW; - *spte = gpte_to_spte(lg, ro_gpte, 0); - } + * 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(lg, 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_u32(lg, gpte_ptr, gpte.raw.val); + lgwrite(lg, gpte_ptr, pte_t, gpte); - /* We succeeded in mapping the page! */ + /* 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; } -/*H:360 (ii) Setting up the page table entry for the Guest stack. +/*H:360 + * (ii) Making sure the Guest stack is mapped. * - * Remember pin_stack_pages() which makes sure the stack is mapped? It could - * simply call demand_page(), but as we've seen that logic is quite long, and - * usually the stack pages are already mapped anyway, so it's not required. + * Remember that direct traps into the Guest need a mapped Guest kernel stack. + * pin_stack_pages() calls us here: we could simply call demand_page(), but as + * we've seen that logic is quite long, and usually the stack pages are already + * 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 lguest *lg, unsigned long vaddr) { - spgd_t *spgd; + pgd_t *spgd; unsigned long flags; - /* Look at the top level entry: is it present? */ + /* Look at the current top level entry: is it present? */ spgd = spgd_addr(lg, lg->pgdidx, vaddr); - if (!(spgd->flags & _PAGE_PRESENT)) + if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) return 0; /* Check the flags on the pte entry itself: it must be present and * writable. */ - flags = spte_addr(lg, *spgd, vaddr)->flags; + flags = pte_flags(*(spte_addr(lg, *spgd, vaddr))); + return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); } @@ -329,39 +324,40 @@ void pin_page(struct lguest *lg, unsigned long vaddr) } /*H:450 If we chase down the release_pgd() code, it looks like this: */ -static void release_pgd(struct lguest *lg, spgd_t *spgd) +static void release_pgd(struct lguest *lg, pgd_t *spgd) { /* If the entry's not present, there's nothing to release. */ - if (spgd->flags & _PAGE_PRESENT) { + if (pgd_flags(*spgd) & _PAGE_PRESENT) { unsigned int i; /* 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). */ - spte_t *ptepage = __va(spgd->pfn << PAGE_SHIFT); + 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 < PTES_PER_PAGE; i++) + 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 PGD entry we we never release it twice. */ - spgd->raw.val = 0; + /* And zero out the PGD entry so we never release it twice. */ + *spgd = __pgd(0); } } -/*H:440 (v) Flushing (thowing away) page tables, - * - * We saw flush_user_mappings() called when we re-used a top-level pgdir page. - * It simply releases every PTE page from 0 up to the kernel address. */ +/*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. */ 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 < vaddr_to_pgd_index(lg->page_offset); i++) + for (i = 0; i < pgd_index(lg->kernel_address); i++) release_pgd(lg, lg->pgdirs[idx].pgdir + i); } -/* The Guest also has a hypercall to do this manually: it's used when a large - * number of mappings have been changed. */ +/*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. */ void guest_pagetable_flush_user(struct lguest *lg) { /* Drop the userspace part of the current page table. */ @@ -369,6 +365,25 @@ void guest_pagetable_flush_user(struct lguest *lg) } /*:*/ +/* We walk down the guest page tables to get a guest-physical address */ +unsigned long guest_pa(struct lguest *lg, unsigned long vaddr) +{ + pgd_t gpgd; + pte_t gpte; + + /* First step: get the top-level Guest page table entry. */ + gpgd = lgread(lg, gpgd_addr(lg, vaddr), pgd_t); + /* Toplevel not present? We can't map it in. */ + if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) + kill_guest(lg, "Bad address %#lx", vaddr); + + gpte = lgread(lg, gpte_addr(lg, gpgd, vaddr), pte_t); + if (!(pte_flags(gpte) & _PAGE_PRESENT)) + kill_guest(lg, "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 * table (if any) corresponding to this top-level address the Guest has given * us. */ @@ -376,7 +391,7 @@ static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) { unsigned int i; for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) - if (lg->pgdirs[i].cr3 == pgtable) + if (lg->pgdirs[i].gpgdir == pgtable) break; return i; } @@ -385,7 +400,7 @@ static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) * allocate a new one (and so the kernel parts are not there), we set * blank_pgdir. */ static unsigned int new_pgdir(struct lguest *lg, - unsigned long cr3, + unsigned long gpgdir, int *blank_pgdir) { unsigned int next; @@ -395,7 +410,7 @@ static unsigned int new_pgdir(struct lguest *lg, next = random32() % ARRAY_SIZE(lg->pgdirs); /* If it's never been allocated at all before, try now. */ if (!lg->pgdirs[next].pgdir) { - lg->pgdirs[next].pgdir = (spgd_t *)get_zeroed_page(GFP_KERNEL); + lg->pgdirs[next].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL); /* If the allocation fails, just keep using the one we have */ if (!lg->pgdirs[next].pgdir) next = lg->pgdidx; @@ -405,7 +420,7 @@ static unsigned int new_pgdir(struct lguest *lg, *blank_pgdir = 1; } /* Record which Guest toplevel this shadows. */ - lg->pgdirs[next].cr3 = cr3; + lg->pgdirs[next].gpgdir = gpgdir; /* Release all the non-kernel mappings. */ flush_user_mappings(lg, next); @@ -414,8 +429,9 @@ static unsigned int new_pgdir(struct lguest *lg, /*H:430 (iv) Switching page tables * - * This is what happens when the Guest changes page tables (ie. changes the - * top-level pgdir). This happens on almost every context switch. */ + * 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 lguest *lg, unsigned long pgtable) { int newpgdir, repin = 0; @@ -434,7 +450,8 @@ void guest_new_pagetable(struct lguest *lg, unsigned long pgtable) } /*H:470 Finally, a routine which throws away everything: all PGD entries in all - * the shadow page tables. This is used when we destroy the Guest. */ + * the shadow page tables, including the Guest's kernel mappings. This is used + * when we destroy the Guest. */ static void release_all_pagetables(struct lguest *lg) { unsigned int i, j; @@ -449,13 +466,22 @@ static void release_all_pagetables(struct lguest *lg) /* 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 is amazingly slow, but thankfully rare. */ + * throw them all away. This traps the Guest in amber for a while as + * everything faults back in, but it's rare. */ void guest_pagetable_clear_all(struct lguest *lg) { release_all_pagetables(lg); /* We need the Guest kernel stack mapped again. */ pin_stack_pages(lg); } +/*:*/ +/*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. :*/ /*H:420 This is the routine which actually sets the page table entry for then * "idx"'th shadow page table. @@ -472,26 +498,28 @@ void guest_pagetable_clear_all(struct lguest *lg) * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately. */ static void do_set_pte(struct lguest *lg, int idx, - unsigned long vaddr, gpte_t gpte) + unsigned long vaddr, pte_t gpte) { - /* Look up the matching shadow page directot entry. */ - spgd_t *spgd = spgd_addr(lg, idx, vaddr); + /* Look up the matching shadow page directory entry. */ + pgd_t *spgd = spgd_addr(lg, idx, vaddr); /* If the top level isn't present, there's no entry to update. */ - if (spgd->flags & _PAGE_PRESENT) { + if (pgd_flags(*spgd) & _PAGE_PRESENT) { /* Otherwise, we start by releasing the existing entry. */ - spte_t *spte = spte_addr(lg, *spgd, vaddr); + pte_t *spte = spte_addr(lg, *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 (gpte.flags & (_PAGE_DIRTY | _PAGE_ACCESSED)) { + if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) { check_gpte(lg, gpte); - *spte = gpte_to_spte(lg, gpte, gpte.flags&_PAGE_DIRTY); + *spte = gpte_to_spte(lg, gpte, + pte_flags(gpte) & _PAGE_DIRTY); } else - /* Otherwise we can demand_page() it in later. */ - spte->raw.val = 0; + /* Otherwise kill it and we can demand_page() it in + * later. */ + *spte = __pte(0); } } @@ -506,18 +534,18 @@ static void do_set_pte(struct lguest *lg, int idx, * The benefit is that when we have to track a new page table, we can copy keep * all the kernel mappings. This speeds up context switch immensely. */ void guest_set_pte(struct lguest *lg, - unsigned long cr3, unsigned long vaddr, gpte_t gpte) + unsigned long gpgdir, unsigned long vaddr, pte_t gpte) { /* Kernel mappings must be changed on all top levels. Slow, but * doesn't happen often. */ - if (vaddr >= lg->page_offset) { + if (vaddr >= lg->kernel_address) { unsigned int i; for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) if (lg->pgdirs[i].pgdir) do_set_pte(lg, i, vaddr, gpte); } else { /* Is this page table one we have a shadow for? */ - int pgdir = find_pgdir(lg, cr3); + int pgdir = find_pgdir(lg, gpgdir); if (pgdir != ARRAY_SIZE(lg->pgdirs)) /* If so, do the update. */ do_set_pte(lg, pgdir, vaddr, gpte); @@ -525,7 +553,7 @@ void guest_set_pte(struct lguest *lg, } /*H:400 - * (iii) Setting up a page table entry when the Guest tells us it has changed. + * (iii) Setting up a page table entry when the Guest tells us one has changed. * * Just like we did in interrupts_and_traps.c, it makes sense for us to deal * with the other side of page tables while we're here: what happens when the @@ -538,7 +566,7 @@ void guest_set_pte(struct lguest *lg, * * So with that in mind here's our code to to update a (top-level) PGD entry: */ -void guest_set_pmd(struct lguest *lg, unsigned long cr3, u32 idx) +void guest_set_pmd(struct lguest *lg, unsigned long gpgdir, u32 idx) { int pgdir; @@ -548,7 +576,7 @@ void guest_set_pmd(struct lguest *lg, unsigned long cr3, u32 idx) return; /* If they're talking about a page table we have a shadow for... */ - pgdir = find_pgdir(lg, cr3); + pgdir = find_pgdir(lg, gpgdir); if (pgdir < ARRAY_SIZE(lg->pgdirs)) /* ... throw it away. */ release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx); @@ -560,21 +588,34 @@ void guest_set_pmd(struct lguest *lg, unsigned long cr3, u32 idx) * its first page table is. We set some things up here: */ int init_guest_pagetable(struct lguest *lg, unsigned long pgtable) { - /* In flush_user_mappings() we loop from 0 to - * "vaddr_to_pgd_index(lg->page_offset)". This assumes it won't hit - * the Switcher mappings, so check that now. */ - if (vaddr_to_pgd_index(lg->page_offset) >= SWITCHER_PGD_INDEX) - return -EINVAL; /* We start on the first shadow page table, and give it a blank PGD * page. */ lg->pgdidx = 0; - lg->pgdirs[lg->pgdidx].cr3 = pgtable; - lg->pgdirs[lg->pgdidx].pgdir = (spgd_t*)get_zeroed_page(GFP_KERNEL); + lg->pgdirs[lg->pgdidx].gpgdir = pgtable; + lg->pgdirs[lg->pgdidx].pgdir = (pgd_t*)get_zeroed_page(GFP_KERNEL); if (!lg->pgdirs[lg->pgdidx].pgdir) return -ENOMEM; return 0; } +/* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ +void page_table_guest_data_init(struct lguest *lg) +{ + /* We get the kernel address: above this is all kernel memory. */ + if (get_user(lg->kernel_address, &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, &lg->lguest_data->reserve_mem) + || put_user(lg->pgdirs[lg->pgdidx].gpgdir,&lg->lguest_data->pgdir)) + kill_guest(lg, "bad guest page %p", lg->lguest_data); + + /* 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(lg->kernel_address) >= SWITCHER_PGD_INDEX) + kill_guest(lg, "bad kernel address %#lx", lg->kernel_address); +} + /* When a Guest dies, our cleanup is fairly simple. */ void free_guest_pagetable(struct lguest *lg) { @@ -589,19 +630,20 @@ void free_guest_pagetable(struct lguest *lg) /*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 available to the + * 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. */ + * 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 lguest *lg, struct lguest_pages *pages) { - spte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages); - spgd_t switcher_pgd; - spte_t regs_pte; + pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages); + pgd_t switcher_pgd; + pte_t regs_pte; /* Make the last PGD entry for this Guest point to the Switcher's PTE * page for this CPU (with appropriate flags). */ - switcher_pgd.pfn = __pa(switcher_pte_page) >> PAGE_SHIFT; - switcher_pgd.flags = _PAGE_KERNEL; + switcher_pgd = __pgd(__pa(switcher_pte_page) | _PAGE_KERNEL); + lg->pgdirs[lg->pgdidx].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; /* We also change the Switcher PTE page. When we're running the Guest, @@ -611,10 +653,8 @@ void map_switcher_in_guest(struct lguest *lg, struct lguest_pages *pages) * 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. */ - regs_pte.pfn = __pa(lg->regs_page) >> PAGE_SHIFT; - regs_pte.flags = _PAGE_KERNEL; - switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTES_PER_PAGE] - = regs_pte; + regs_pte = pfn_pte (__pa(lg->regs_page) >> PAGE_SHIFT, __pgprot(_PAGE_KERNEL)); + switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTRS_PER_PTE] = regs_pte; } /*:*/ @@ -635,26 +675,39 @@ static __init void populate_switcher_pte_page(unsigned int cpu, unsigned int pages) { unsigned int i; - spte_t *pte = switcher_pte_page(cpu); + pte_t *pte = switcher_pte_page(cpu); /* The first entries are easy: they map the Switcher code. */ for (i = 0; i < pages; i++) { - pte[i].pfn = page_to_pfn(switcher_page[i]); - pte[i].flags = _PAGE_PRESENT|_PAGE_ACCESSED; + pte[i] = mk_pte(switcher_page[i], + __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)); } /* 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 = page_to_pfn(switcher_page[i]); - pte[i].flags = _PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW; + 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 = page_to_pfn(switcher_page[i+1]); - pte[i+1].flags = _PAGE_PRESENT|_PAGE_ACCESSED; + pte[i+1] = pfn_pte(page_to_pfn(switcher_page[i+1]), + __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)); } +/* 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 in sync with the Guest's page tables is for one reason: for most + * Guests this page table dance determines how bad performance will be. This + * is why Xen 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) @@ -662,7 +715,7 @@ __init int init_pagetables(struct page **switcher_page, unsigned int pages) unsigned int i; for_each_possible_cpu(i) { - switcher_pte_page(i) = (spte_t *)get_zeroed_page(GFP_KERNEL); + switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL); if (!switcher_pte_page(i)) { free_switcher_pte_pages(); return -ENOMEM;