* 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 to point the hardware to the
- * actual Guest pages when running the Guest. :*/
+ * it nor use it: we verify and convert it here then point the CPU to the
+ * converted Guest pages when running the Guest. :*/
/* Copyright (C) Rusty Russell IBM Corporation 2006.
* GPL v2 and any later version */
#include <linux/percpu.h>
#include <asm/tlbflush.h>
#include <asm/uaccess.h>
+#include <asm/bootparam.h>
#include "lg.h"
/*M:008 We hold reference to pages, which prevents them from being swapped.
*
* 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
*
* 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.
:*/
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.
*
* 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 pgd_t *spgd_addr(struct lguest *lg, u32 i, unsigned long vaddr)
+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(lg, "attempt to access switcher pages");
+ kill_guest(cpu, "attempt to access switcher pages");
index = 0;
}
/* Return a pointer index'th pgd entry for the i'th page table. */
- return &lg->pgdirs[i].pgdir[index];
+ return &cpu->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 pte_t *spte_addr(struct lguest *lg, pgd_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(pgd_t spgd, unsigned long vaddr)
{
pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT);
/* You should never call this if the PGD entry wasn't valid */
/* 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)
+static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr)
{
unsigned int index = vaddr >> (PGDIR_SHIFT);
- return lg->pgdirs[lg->pgdidx].gpgdir + index * sizeof(pgd_t);
+ return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t);
}
-static unsigned long gpte_addr(struct lguest *lg,
- pgd_t gpgd, unsigned long vaddr)
+static unsigned long gpte_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);
}
+/*:*/
+
+/*M:014 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
* an actual, physical page number. It can fail for several reasons: the
* 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. */
+ * This holds a reference to the page, so release_pte() is careful to put that
+ * 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
* 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 pte_t gpte_to_spte(struct lguest *lg, pte_t gpte, int write)
+static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write)
{
unsigned long pfn, base, flags;
flags = (pte_flags(gpte) & ~_PAGE_GLOBAL);
/* The Guest's pages are offset inside the Launcher. */
- base = (unsigned long)lg->mem_base / PAGE_SIZE;
+ base = (unsigned long)cpu->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
* page, given the virtual number. */
pfn = get_pfn(base + pte_pfn(gpte), write);
if (pfn == -1UL) {
- kill_guest(lg, "failed to get page %lu", pte_pfn(gpte));
+ kill_guest(cpu, "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! */
/*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
+ /* 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)));
}
/*:*/
-static void check_gpte(struct lguest *lg, pte_t gpte)
+static void check_gpte(struct lg_cpu *cpu, pte_t gpte)
{
- if ((pte_flags(gpte) & (_PAGE_PWT|_PAGE_PSE))
- || pte_pfn(gpte) >= lg->pfn_limit)
- kill_guest(lg, "bad page table entry");
+ if ((pte_flags(gpte) & _PAGE_PSE) ||
+ pte_pfn(gpte) >= cpu->lg->pfn_limit)
+ kill_guest(cpu, "bad page table entry");
}
-static void check_gpgd(struct lguest *lg, pgd_t gpgd)
+static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd)
{
- if ((pgd_flags(gpgd) & ~_PAGE_TABLE) || pgd_pfn(gpgd) >= lg->pfn_limit)
- kill_guest(lg, "bad page directory entry");
+ if ((pgd_flags(gpgd) & ~_PAGE_TABLE) ||
+ (pgd_pfn(gpgd) >= cpu->lg->pfn_limit))
+ kill_guest(cpu, "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
* and return to the Guest without it knowing.
*
* If we fixed up the fault (ie. we mapped the address), this routine returns
- * true. */
-int demand_page(struct lguest *lg, 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;
pte_t *spte;
/* First step: get the top-level Guest page table entry. */
- gpgd = lgread(lg, gpgd_addr(lg, vaddr), pgd_t);
+ 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;
+ return false;
/* Now look at the matching shadow entry. */
- spgd = spgd_addr(lg, lg->pgdidx, vaddr);
+ 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(lg, "out of memory allocating pte page");
- return 0;
+ kill_guest(cpu, "out of memory allocating pte page");
+ return false;
}
/* We check that the Guest pgd is OK. */
- check_gpgd(lg, gpgd);
+ 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));
/* 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 = lgread(lg, gpte_ptr, pte_t);
+ gpte_ptr = gpte_addr(gpgd, vaddr);
+ 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). */
if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW))
- return 0;
+ return false;
- /* User access to a kernel page? (bit 3 == user access) */
+ /* 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(lg, gpte);
+ check_gpte(cpu, gpte);
+
/* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
gpte = pte_mkyoung(gpte);
-
if (errcode & 2)
gpte = pte_mkdirty(gpte);
/* Get the pointer to the shadow PTE entry we're going to set. */
- spte = spte_addr(lg, *spgd, vaddr);
+ 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. */
release_pte(*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(lg, gpte, 1);
+ *spte = gpte_to_spte(cpu, gpte, 1);
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. */
- *spte = gpte_to_spte(lg, pte_wrprotect(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(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(lg, gpte_ptr, pte_t, gpte);
+ lgwrite(cpu, gpte_ptr, pte_t, gpte);
- /* We succeeded in mapping the page! */
- return 1;
+ /* 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 true;
}
-/*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)
+static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr)
{
pgd_t *spgd;
unsigned long flags;
- /* Look at the top level entry: is it present? */
- spgd = spgd_addr(lg, lg->pgdidx, vaddr);
+ /* 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;
+ return false;
/* Check the flags on the pte entry itself: it must be present and
* writable. */
- flags = pte_flags(*(spte_addr(lg, *spgd, vaddr)));
+ flags = pte_flags(*(spte_addr(*spgd, vaddr)));
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
* in the page tables, and if not, we call demand_page() with error code 2
* (meaning "write"). */
-void pin_page(struct lguest *lg, unsigned long vaddr)
+void pin_page(struct lg_cpu *cpu, unsigned long vaddr)
{
- if (!page_writable(lg, vaddr) && !demand_page(lg, vaddr, 2))
- kill_guest(lg, "bad stack page %#lx", vaddr);
+ if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2))
+ kill_guest(cpu, "bad stack page %#lx", vaddr);
}
/*H:450 If we chase down the release_pgd() code, it looks like this: */
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. */
+ /* 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_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. */
-void guest_pagetable_flush_user(struct lguest *lg)
+/*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 lg_cpu *cpu)
{
/* Drop the userspace part of the current page table. */
- flush_user_mappings(lg, lg->pgdidx);
+ flush_user_mappings(cpu->lg, cpu->cpu_pgd);
}
/*:*/
/* We walk down the guest page tables to get a guest-physical address */
-unsigned long guest_pa(struct lguest *lg, unsigned long vaddr)
+unsigned long guest_pa(struct lg_cpu *cpu, 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);
+ gpgd = lgread(cpu, gpgd_addr(cpu, 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);
+ if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) {
+ kill_guest(cpu, "Bad address %#lx", vaddr);
+ return -1UL;
+ }
- gpte = lgread(lg, gpte_addr(lg, gpgd, vaddr), pte_t);
+ gpte = lgread(cpu, gpte_addr(gpgd, vaddr), pte_t);
if (!(pte_flags(gpte) & _PAGE_PRESENT))
- kill_guest(lg, "Bad address %#lx", vaddr);
+ kill_guest(cpu, "Bad address %#lx", vaddr);
return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK);
}
{
unsigned int i;
for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
- if (lg->pgdirs[i].gpgdir == pgtable)
+ if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable)
break;
return i;
}
/*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. */
-static unsigned int new_pgdir(struct lguest *lg,
+static unsigned int new_pgdir(struct lg_cpu *cpu,
unsigned long gpgdir,
int *blank_pgdir)
{
/* 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(lg->pgdirs);
+ next = random32() % ARRAY_SIZE(cpu->lg->pgdirs);
/* If it's never been allocated at all before, try now. */
- if (!lg->pgdirs[next].pgdir) {
- lg->pgdirs[next].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL);
+ if (!cpu->lg->pgdirs[next].pgdir) {
+ cpu->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;
+ 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! */
*blank_pgdir = 1;
}
/* Record which Guest toplevel this shadows. */
- lg->pgdirs[next].gpgdir = gpgdir;
+ cpu->lg->pgdirs[next].gpgdir = gpgdir;
/* Release all the non-kernel mappings. */
- flush_user_mappings(lg, next);
+ flush_user_mappings(cpu->lg, next);
return next;
}
/*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. */
-void guest_new_pagetable(struct lguest *lg, unsigned long pgtable)
+ * 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)
{
int newpgdir, repin = 0;
/* Look to see if we have this one already. */
- newpgdir = find_pgdir(lg, pgtable);
+ 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(lg->pgdirs))
- newpgdir = new_pgdir(lg, pgtable, &repin);
+ if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs))
+ newpgdir = new_pgdir(cpu, pgtable, &repin);
/* Change the current pgd index to the new one. */
- lg->pgdidx = newpgdir;
+ cpu->cpu_pgd = newpgdir;
/* If it was completely blank, we map in the Guest kernel stack */
if (repin)
- pin_stack_pages(lg);
+ pin_stack_pages(cpu);
}
/*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;
/* 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. */
-void guest_pagetable_clear_all(struct lguest *lg)
+ * 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 lg_cpu *cpu)
{
- release_all_pagetables(lg);
+ release_all_pagetables(cpu->lg);
/* We need the Guest kernel stack mapped again. */
- pin_stack_pages(lg);
+ pin_stack_pages(cpu);
}
+/*:*/
+/*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.
* _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 lguest *lg, int idx,
+static void do_set_pte(struct lg_cpu *cpu, int idx,
unsigned long vaddr, pte_t gpte)
{
- /* Look up the matching shadow page directot entry. */
- pgd_t *spgd = spgd_addr(lg, idx, vaddr);
+ /* Look up the matching shadow page directory entry. */
+ pgd_t *spgd = spgd_addr(cpu, idx, vaddr);
/* 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(lg, *spgd, vaddr);
+ 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(lg, gpte);
- *spte = gpte_to_spte(lg, gpte,
+ check_gpte(cpu, gpte);
+ *spte = gpte_to_spte(cpu, gpte,
pte_flags(gpte) & _PAGE_DIRTY);
} else
- /* Otherwise we can demand_page() it in later. */
+ /* Otherwise kill it and we can demand_page() it in
+ * later. */
*spte = __pte(0);
}
}
* all processes. So when the page table above that address changes, we update
* 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 copy keep
- * all the kernel mappings. This speeds up context switch immensely. */
-void guest_set_pte(struct lguest *lg,
+ * 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. */
+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. */
- if (vaddr >= lg->kernel_address) {
+ /* 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(lg->pgdirs); i++)
- if (lg->pgdirs[i].pgdir)
- do_set_pte(lg, i, vaddr, gpte);
+ for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++)
+ if (cpu->lg->pgdirs[i].pgdir)
+ do_set_pte(cpu, i, vaddr, gpte);
} else {
/* Is this page table one we have a shadow for? */
- int pgdir = find_pgdir(lg, gpgdir);
- if (pgdir != ARRAY_SIZE(lg->pgdirs))
+ int pgdir = find_pgdir(cpu->lg, gpgdir);
+ if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs))
/* If so, do the update. */
- do_set_pte(lg, pgdir, vaddr, gpte);
+ do_set_pte(cpu, pgdir, vaddr, gpte);
}
}
/*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
release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx);
}
+/* Once we know how much memory we have we can construct simple identity
+ * (which set virtual == physical) and linear mappings
+ * which will get the Guest far enough into the boot to create its own.
+ *
+ * We lay them out of the way, just below the initrd (which is why we need to
+ * know its size here). */
+static unsigned long setup_pagetables(struct lguest *lg,
+ unsigned long mem,
+ unsigned long initrd_size)
+{
+ pgd_t __user *pgdir;
+ pte_t __user *linear;
+ unsigned int mapped_pages, i, linear_pages, phys_linear;
+ unsigned long mem_base = (unsigned long)lg->mem_base;
+
+ /* We have mapped_pages frames to map, so we need
+ * linear_pages page tables to map them. */
+ mapped_pages = mem / PAGE_SIZE;
+ linear_pages = (mapped_pages + PTRS_PER_PTE - 1) / PTRS_PER_PTE;
+
+ /* We put the toplevel page directory page at the top of memory. */
+ pgdir = (pgd_t *)(mem + mem_base - initrd_size - PAGE_SIZE);
+
+ /* Now we use the next linear_pages pages as pte pages */
+ linear = (void *)pgdir - linear_pages * PAGE_SIZE;
+
+ /* Linear mapping is easy: put every page's address into the
+ * mapping in order. */
+ for (i = 0; i < mapped_pages; i++) {
+ pte_t pte;
+ pte = pfn_pte(i, __pgprot(_PAGE_PRESENT|_PAGE_RW|_PAGE_USER));
+ if (copy_to_user(&linear[i], &pte, sizeof(pte)) != 0)
+ return -EFAULT;
+ }
+
+ /* The top level points to the linear page table pages above.
+ * We setup the identity and linear mappings here. */
+ phys_linear = (unsigned long)linear - mem_base;
+ for (i = 0; i < mapped_pages; i += PTRS_PER_PTE) {
+ pgd_t pgd;
+ pgd = __pgd((phys_linear + i * sizeof(pte_t)) |
+ (_PAGE_PRESENT | _PAGE_RW | _PAGE_USER));
+
+ if (copy_to_user(&pgdir[i / PTRS_PER_PTE], &pgd, sizeof(pgd))
+ || copy_to_user(&pgdir[pgd_index(PAGE_OFFSET)
+ + i / PTRS_PER_PTE],
+ &pgd, sizeof(pgd)))
+ return -EFAULT;
+ }
+
+ /* We return the top level (guest-physical) address: remember where
+ * this is. */
+ return (unsigned long)pgdir - mem_base;
+}
+
/*H:500 (vii) Setting up the page tables initially.
*
* 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)
+int init_guest_pagetable(struct lguest *lg)
{
+ u64 mem;
+ u32 initrd_size;
+ struct boot_params __user *boot = (struct boot_params *)lg->mem_base;
+
+ /* Get the Guest memory size and the ramdisk size from the boot header
+ * located at lg->mem_base (Guest address 0). */
+ if (copy_from_user(&mem, &boot->e820_map[0].size, sizeof(mem))
+ || get_user(initrd_size, &boot->hdr.ramdisk_size))
+ return -EFAULT;
+
/* We start on the first shadow page table, and give it a blank PGD
* page. */
- lg->pgdidx = 0;
- lg->pgdirs[lg->pgdidx].gpgdir = pgtable;
- lg->pgdirs[lg->pgdidx].pgdir = (pgd_t*)get_zeroed_page(GFP_KERNEL);
- if (!lg->pgdirs[lg->pgdidx].pgdir)
+ lg->pgdirs[0].gpgdir = setup_pagetables(lg, mem, initrd_size);
+ if (IS_ERR_VALUE(lg->pgdirs[0].gpgdir))
+ return lg->pgdirs[0].gpgdir;
+ lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL);
+ if (!lg->pgdirs[0].pgdir)
return -ENOMEM;
+ lg->cpus[0].cpu_pgd = 0;
return 0;
}
/* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
-void page_table_guest_data_init(struct lguest *lg)
+void page_table_guest_data_init(struct lg_cpu *cpu)
{
/* We get the kernel address: above this is all kernel memory. */
- if (get_user(lg->kernel_address, &lg->lguest_data->kernel_address)
+ 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, &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);
+ || put_user(4U*1024*1024, &cpu->lg->lguest_data->reserve_mem)
+ || put_user(cpu->lg->pgdirs[0].gpgdir, &cpu->lg->lguest_data->pgdir))
+ kill_guest(cpu, "bad guest page %p", cpu->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);
+ if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX)
+ kill_guest(cpu, "bad kernel address %#lx",
+ cpu->lg->kernel_address);
}
/* When a Guest dies, our cleanup is fairly simple. */
/*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. */
-void map_switcher_in_guest(struct lguest *lg, struct lguest_pages *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)
{
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);
+ switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL);
- lg->pgdirs[lg->pgdidx].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd;
+ cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd;
/* 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
* 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_pte (__pa(lg->regs_page) >> PAGE_SHIFT, __pgprot(_PAGE_KERNEL));
+ 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;
}
/*:*/
__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)
git://ftp.safe.ca / safe / jmp / linux-2.6 / blobdiff