[Lguest] NULL pointer dereference at __switch_to() ( __unlazy_fpu ) with lguest PAE patch
Matias Zabaljauregui
zabaljauregui at gmail.com
Fri Apr 17 03:21:14 EST 2009
Hi,
For some days I have been looking for the bug that causes an easily reproducible oops in the guest
when I apply my PAE support _draft_ patch (appended at the end of this mail) to lguest.
This is the oops:
Setting kernel variables...done.
Will now mount local filesystems:.
Will now activate swapfile swap:done.
Cleaning /tmp...
[ 84.749676] BUG: unable to handle kernel NULL pointer dereference at 00000005
[ 84.749676] IP: [<c0101f6e>] __switch_to+0xd/0x12d
[ 84.749676] *pdpt = 000000001fa12001 *pde = 0000000000000000
[ 84.749676] Oops: 0000 [#1] PREEMPT
[ 84.749676] last sysfs file: /sys/kernel/uevent_seqnum
[ 84.749676] Modules linked in:
[ 84.749676]
[ 84.749676] Pid: 1066, comm: find Not tainted (2.6.30-rc2-00167-gcd97824-dirty #1)
[ 84.749676] EIP: 0061:[<c0101f6e>] EFLAGS: 00000092 CPU: 0
[ 84.749676] EIP is at __switch_to+0xd/0x12d
[ 84.749676] EAX: 00000001 EBX: dfa371b0 ECX: df8b0430 EDX: dfa371b0
[ 84.749676] ESI: 00000001 EDI: df887200 EBP: df865ec4 ESP: df865eac
[ 84.749676] DS: 007b ES: 007b FS: 0000 GS: 0000 SS: 0069
[ 84.749676] Process find (pid: 1066, ti=df864000 task=df8b0430 task.ti=dfa0e000)
[ 84.749676] Stack:
[ 84.749676] 00000000 00000001 df8b0464 dfa371b0 df8b0430 df887200 df865ee0 c0101b7d
[ 84.749676] 00000004 c040f544 dfa371b0 dfa13bc0 dfa13540 dfa0ff58 c03211b7 df865f28
[ 84.749676] 00000286 00000000 00393bc7 df865f20 dfa371b0 dfa37340 dfa5d8a0 dfa371b0
[ 84.749676] Call Trace:
[ 84.749676] [<c0101b7d>] ? lazy_hcall1+0x32/0xac
[ 84.749676] [<c03211b7>] ? __schedule+0x2c2/0x31f
[ 84.749676] [<c0321226>] ? schedule+0x12/0x24
[ 84.749676] [<c01225ff>] ? do_wait+0x1ec/0x363
[ 84.749676] [<c011c4a7>] ? default_wake_function+0x0/0xd
[ 84.749676] [<c020fabe>] ? copy_to_user+0x2a/0x34
[ 84.749676] [<c01227e5>] ? sys_wait4+0x6f/0x85
[ 84.749676] [<c012280e>] ? sys_waitpid+0x13/0x15
[ 84.749676] [<c01037c5>] ? syscall_call+0x7/0xb
[ 84.749676] Code: 00 01 80 00 6a 00 6a 00 6a 00 8d 4d b0 31 d2 89 f0 e8 d3 d7 01 00 8d 65 f4 5b 5e 5f c9 c3 55 89 e5 57 56 53 83 ec 0c 89 c6 89 d3 <8b> 40 04 8b 40 0c a8 01 74 56 a8 10 8b be 60 02 00 00 74 1b 83
[ 84.749676] EIP: [<c0101f6e>] __switch_to+0xd/0x12d SS:ESP 0069:df865eac
[ 84.749676] CR2: 0000000000000005
[ 84.749676] ---[ end trace 54cfaaa2a7bf67ca ]---
[ 84.749676] Fixing recursive fault but reboot is needed!
and looking for the NULL dereference, it seems to be in __unlazy_fpu
# gdb -q vmlinux
(gdb) list *0xc0101f6e
0xc0101f6e is in __switch_to (/usr/src/linux-2.6/arch/x86/include/asm/i387.h:273).
268 extern int save_i387_xstate(void __user *buf);
269 extern int restore_i387_xstate(void __user *buf);
270
271 static inline void __unlazy_fpu(struct task_struct *tsk)
272 {
273 if (task_thread_info(tsk)->status & TS_USEDFPU) {
274 __save_init_fpu(tsk);
275 stts();
276 } else
277 tsk->fpu_counter = 0;
This oops disappears when I use no387 and nofxsr guest kernel parameters in lguest command invocation
Now, this is only happening with my PAE patch applied, so I assume that my code is broken.
But these seems to be the same symptoms discussed in this thread:
http://lkml.indiana.edu/hypermail/linux/kernel/0806.2/0787.html
So I thought that maybe you can help me with some hints.
I really appreciate your help,
Matias
Here is my patch:
diff --git a/arch/x86/include/asm/lguest.h b/arch/x86/include/asm/lguest.h
index 1caf576..ffbf1ac 100644
--- a/arch/x86/include/asm/lguest.h
+++ b/arch/x86/include/asm/lguest.h
@@ -17,8 +17,13 @@
/* Pages for switcher itself, then two pages per cpu */
#define TOTAL_SWITCHER_PAGES (SHARED_SWITCHER_PAGES + 2 * nr_cpu_ids)
+#ifdef CONFIG_X86_PAE
+/* We map at -2M for ease of mapping into the guest (one PTE page). */
+#define SWITCHER_ADDR 0xFFE00000
+#else
/* We map at -4M for ease of mapping into the guest (one PTE page). */
#define SWITCHER_ADDR 0xFFC00000
+#endif
/* Found in switcher.S */
extern unsigned long default_idt_entries[];
diff --git a/arch/x86/include/asm/lguest_hcall.h b/arch/x86/include/asm/lguest_hcall.h
index 0f4ee71..3860153 100644
--- a/arch/x86/include/asm/lguest_hcall.h
+++ b/arch/x86/include/asm/lguest_hcall.h
@@ -17,6 +17,7 @@
#define LHCALL_SET_PMD 15
#define LHCALL_LOAD_TLS 16
#define LHCALL_NOTIFY 17
+#define LHCALL_SET_PUD 18
#define LGUEST_TRAP_ENTRY 0x1F
diff --git a/arch/x86/lguest/Kconfig b/arch/x86/lguest/Kconfig
index 8dab8f7..3871804 100644
--- a/arch/x86/lguest/Kconfig
+++ b/arch/x86/lguest/Kconfig
@@ -2,7 +2,6 @@ config LGUEST_GUEST
bool "Lguest guest support"
select PARAVIRT
depends on X86_32
- depends on !X86_PAE
select VIRTIO
select VIRTIO_RING
select VIRTIO_CONSOLE
diff --git a/arch/x86/lguest/boot.c b/arch/x86/lguest/boot.c
index e94a11e..ce7b010 100644
--- a/arch/x86/lguest/boot.c
+++ b/arch/x86/lguest/boot.c
@@ -359,8 +359,12 @@ static void lguest_cpuid(unsigned int *ax, unsigned int *bx,
case 1: /* Basic feature request. */
/* We only allow kernel to see SSE3, CMPXCHG16B and SSSE3 */
*cx &= 0x00002201;
- /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU. */
+ /* SSE, SSE2, FXSR, MMX, CMOV, CMPXCHG8B, TSC, FPU, PAE. */
+#ifdef CONFIG_X86_PAE
+ *dx &= 0x07808151;
+#else
*dx &= 0x07808111;
+#endif
/* The Host can do a nice optimization if it knows that the
* kernel mappings (addresses above 0xC0000000 or whatever
* PAGE_OFFSET is set to) haven't changed. But Linux calls
@@ -518,18 +522,30 @@ static void lguest_pte_update(struct mm_struct *mm, unsigned long addr,
static void lguest_set_pte_at(struct mm_struct *mm, unsigned long addr,
pte_t *ptep, pte_t pteval)
{
- *ptep = pteval;
+ native_set_pte(ptep, pteval);
lguest_pte_update(mm, addr, ptep);
}
+#ifdef CONFIG_X86_PAE
/* The Guest calls this to set a top-level entry. Again, we set the entry then
* tell the Host which top-level page we changed, and the index of the entry we
* changed. */
+static void lguest_set_pud(pud_t *pudp, pud_t pudval)
+{
+ native_set_pud (pudp, pudval);
+
+ /* 32 bytes aligned pdpt address. */
+ lazy_hcall2(LHCALL_SET_PUD, __pa(pudp) & 0xFFFFFFE0,
+ (__pa(pudp) & 0x1F) / sizeof(pud_t));
+}
+#endif
+
+/* The Guest calls this to set a PMD entry, when PAE is active */
static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
{
- *pmdp = pmdval;
+ native_set_pmd (pmdp, pmdval);
lazy_hcall2(LHCALL_SET_PMD, __pa(pmdp) & PAGE_MASK,
- (__pa(pmdp) & (PAGE_SIZE - 1)) / 4);
+ (__pa(pmdp) & (PAGE_SIZE - 1)) / sizeof(pmd_t));
}
/* There are a couple of legacy places where the kernel sets a PTE, but we
@@ -543,11 +559,31 @@ static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
* which brings boot back to 0.25 seconds. */
static void lguest_set_pte(pte_t *ptep, pte_t pteval)
{
- *ptep = pteval;
+ native_set_pte(ptep, pteval);
+ if (cr3_changed)
+ lazy_hcall1(LHCALL_FLUSH_TLB, 1);
+}
+
+#ifdef CONFIG_X86_PAE
+static void lguest_set_pte_atomic(pte_t *ptep, pte_t pte)
+{
+ native_set_pte_atomic(ptep, pte);
if (cr3_changed)
lazy_hcall1(LHCALL_FLUSH_TLB, 1);
}
+void lguest_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
+{
+ native_pte_clear(mm, addr, ptep);
+ lazy_hcall3(LHCALL_SET_PTE, lguest_data.pgdir, addr, 0);
+}
+
+void lguest_pmd_clear(pmd_t *pmdp)
+{
+ lguest_set_pmd(pmdp, __pmd(0));
+}
+#endif
+
/* Unfortunately for Lguest, the pv_mmu_ops for page tables were based on
* native page table operations. On native hardware you can set a new page
* table entry whenever you want, but if you want to remove one you have to do
@@ -1017,6 +1053,7 @@ __init void lguest_init(void)
pv_info.name = "lguest";
pv_info.paravirt_enabled = 1;
pv_info.kernel_rpl = 1;
+ pv_info.shared_kernel_pmd = 1;
/* We set up all the lguest overrides for sensitive operations. These
* are detailed with the operations themselves. */
@@ -1062,6 +1099,13 @@ __init void lguest_init(void)
pv_mmu_ops.set_pte = lguest_set_pte;
pv_mmu_ops.set_pte_at = lguest_set_pte_at;
pv_mmu_ops.set_pmd = lguest_set_pmd;
+
+#ifdef CONFIG_X86_PAE
+ pv_mmu_ops.set_pte_atomic = lguest_set_pte_atomic;
+ pv_mmu_ops.pte_clear = lguest_pte_clear;
+ pv_mmu_ops.pmd_clear = lguest_pmd_clear;
+ pv_mmu_ops.set_pud = lguest_set_pud;
+#endif
pv_mmu_ops.read_cr2 = lguest_read_cr2;
pv_mmu_ops.read_cr3 = lguest_read_cr3;
pv_mmu_ops.lazy_mode.enter = paravirt_enter_lazy_mmu;
diff --git a/drivers/lguest/Kconfig b/drivers/lguest/Kconfig
index a3d3cba..8f63845 100644
--- a/drivers/lguest/Kconfig
+++ b/drivers/lguest/Kconfig
@@ -1,6 +1,6 @@
config LGUEST
tristate "Linux hypervisor example code"
- depends on X86_32 && EXPERIMENTAL && !X86_PAE && FUTEX
+ depends on X86_32 && EXPERIMENTAL && FUTEX
select HVC_DRIVER
---help---
This is a very simple module which allows you to run
diff --git a/drivers/lguest/hypercalls.c b/drivers/lguest/hypercalls.c
index 54d66f0..c5d6678 100644
--- a/drivers/lguest/hypercalls.c
+++ b/drivers/lguest/hypercalls.c
@@ -78,6 +78,11 @@ static void do_hcall(struct lg_cpu *cpu, struct hcall_args *args)
case LHCALL_SET_PMD:
guest_set_pmd(cpu->lg, args->arg1, args->arg2);
break;
+#ifdef CONFIG_X86_PAE
+ case LHCALL_SET_PUD:
+ guest_set_pud(cpu->lg, args->arg1, args->arg2);
+ break;
+#endif
case LHCALL_SET_CLOCKEVENT:
guest_set_clockevent(cpu, args->arg1);
break;
diff --git a/drivers/lguest/lg.h b/drivers/lguest/lg.h
index ac8a4a3..514a6c0 100644
--- a/drivers/lguest/lg.h
+++ b/drivers/lguest/lg.h
@@ -18,7 +18,7 @@ int init_pagetables(struct page **switcher_page, unsigned int pages);
struct pgdir
{
- unsigned long gpgdir;
+ pgd_t *gpgdir;
pgd_t *pgdir;
};
@@ -137,6 +137,8 @@ int run_guest(struct lg_cpu *cpu, unsigned long __user *user);
* in the kernel. */
#define pgd_flags(x) (pgd_val(x) & ~PAGE_MASK)
#define pgd_pfn(x) (pgd_val(x) >> PAGE_SHIFT)
+#define pmd_flags(x) (pmd_val(x) & ~PAGE_MASK)
+#define pmd_pfn(x) (pmd_val(x) >> PAGE_SHIFT)
/* interrupts_and_traps.c: */
void maybe_do_interrupt(struct lg_cpu *cpu);
@@ -168,6 +170,9 @@ int init_guest_pagetable(struct lguest *lg);
void free_guest_pagetable(struct lguest *lg);
void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable);
void guest_set_pmd(struct lguest *lg, unsigned long gpgdir, u32 i);
+#ifdef CONFIG_X86_PAE
+void guest_set_pud(struct lguest *lg, unsigned long gpgdir, u32 i);
+#endif
void guest_pagetable_clear_all(struct lg_cpu *cpu);
void guest_pagetable_flush_user(struct lg_cpu *cpu);
void guest_set_pte(struct lg_cpu *cpu, unsigned long gpgdir,
diff --git a/drivers/lguest/page_tables.c b/drivers/lguest/page_tables.c
index a059cf9..77014d8 100644
--- a/drivers/lguest/page_tables.c
+++ b/drivers/lguest/page_tables.c
@@ -47,12 +47,20 @@
* (vii) Setting up the page tables initially.
:*/
+void guest_pagetable_clear_all(struct lg_cpu *cpu);
/* 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 (PTRS_PER_PGD - 1)
+/* For PAE we need the PMD index as well. We can use the last 2MB, so we
+ * will need the last pmd entry of the last pmd page. */
+#ifdef CONFIG_X86_PAE
+
+#define SWITCHER_PMD_INDEX (PTRS_PER_PMD - 1)
+#endif
+
/* 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. */
@@ -73,39 +81,90 @@ static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr)
{
unsigned int index = pgd_index(vaddr);
+#ifndef CONFIG_X86_PAE
/* 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;
}
+#endif
/* Return a pointer index'th pgd entry for the i'th page table. */
return &cpu->lg->pgdirs[i].pgdir[index];
}
+#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;
+
+ /* We kill any Guest trying to touch the Switcher addresses. */
+ if (pgd_index(vaddr) == SWITCHER_PGD_INDEX &&
+ index >= SWITCHER_PMD_INDEX) {
+ kill_guest(cpu, "attempt to access switcher pages");
+ index = 0;
+ }
+
+ /* 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)
+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. */
-static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr)
+static pgd_t *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);
+ return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index;
+}
+
+#ifdef CONFIG_X86_PAE
+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 + pmd_index(vaddr) * sizeof(pmd_t);
}
+#endif
-static unsigned long gpte_addr(pgd_t gpgd, unsigned long vaddr)
+static unsigned long gpte_addr(struct lg_cpu *cpu,
+ pgd_t gpgd, unsigned long vaddr)
{
+#ifdef CONFIG_X86_PAE
+ pmd_t gpmd = lgread(cpu,
+ (unsigned long) gpmd_addr(gpgd, vaddr), pmd_t);
+ unsigned long gpage = pmd_pfn(gpmd) << PAGE_SHIFT;
+#else
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);
+#endif
+ return gpage + pte_index(vaddr) * sizeof(pte_t);
}
/*:*/
@@ -184,11 +243,24 @@ static void check_gpte(struct lg_cpu *cpu, pte_t gpte)
static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd)
{
+#ifdef CONFIG_X86_PAE
+ if ((pgd_flags(gpgd) & ~_PAGE_PRESENT) ||
+#else
if ((pgd_flags(gpgd) & ~_PAGE_TABLE) ||
+#endif
(pgd_pfn(gpgd) >= cpu->lg->pfn_limit))
kill_guest(cpu, "bad page directory entry");
}
+#ifdef CONFIG_X86_PAE
+static void 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");
+}
+#endif
+
/*H:330
* (i) Looking up a page table entry when the Guest faults.
*
@@ -207,8 +279,14 @@ bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode)
pte_t gpte;
pte_t *spte;
+#ifdef CONFIG_X86_PAE
+ pmd_t *spmd;
+ pmd_t gpmd;
+#endif
+
/* First step: get the top-level Guest page table entry. */
- gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
+ gpgd = lgread(cpu, (unsigned long) gpgd_addr(cpu, vaddr), pgd_t);
+
/* Toplevel not present? We can't map it in. */
if (!(pgd_flags(gpgd) & _PAGE_PRESENT))
return false;
@@ -231,9 +309,38 @@ bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode)
*spgd = __pgd(__pa(ptepage) | pgd_flags(gpgd));
}
+#ifdef CONFIG_X86_PAE
+ gpmd = lgread(cpu, (unsigned long) gpmd_addr(gpgd, vaddr), pmd_t);
+ /* middle level not present? We can't map it in. */
+ if (!(pmd_flags(gpmd) & _PAGE_PRESENT))
+ return 0;
+
+ /* Now look at the matching 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 = 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 pmd is OK. */
+ check_gpmd(cpu, gpmd);
+
+ /* And we copy the flags to the shadow PMD entry. The page
+ * number in the shadow PMD is the page we just allocated. */
+ *spmd = __pmd(__pa(ptepage) | pmd_flags(gpmd));
+ }
+#endif
+
/* 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_ptr = gpte_addr(cpu, 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. */
@@ -259,7 +366,7 @@ bool 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);
+ spte = spte_addr(cpu, *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);
@@ -301,14 +408,24 @@ static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr)
pgd_t *spgd;
unsigned long flags;
+#ifdef CONFIG_X86_PAE
+ pmd_t *spmd;
+#endif
+
/* 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 false;
+#ifdef CONFIG_X86_PAE
+ spmd = spmd_addr(cpu, *spgd, vaddr);
+ if (!(pmd_flags(*spmd) & _PAGE_PRESENT))
+ return false;
+#endif
+
/* Check the flags on the pte entry itself: it must be present and
* writable. */
- flags = pte_flags(*(spte_addr(*spgd, vaddr)));
+ flags = pte_flags(*(spte_addr(cpu, *spgd, vaddr)));
return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW);
}
@@ -322,8 +439,45 @@ void pin_page(struct lg_cpu *cpu, unsigned long vaddr)
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. */
+ native_set_pmd(spmd, __pmd(0));
+ }
+}
+
+/*H:450 If we chase down the release_pgd() code, it looks 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;
+ 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. */
+ native_set_pud ((pud_t *)spgd, __pud(0));
+ }
+}
+
+#else /* !CONFIG_X86_PAE */
+
/*H:450 If we chase down the release_pgd() code, it looks like this: */
-static void release_pgd(struct lguest *lg, pgd_t *spgd)
+static void release_pgd(pgd_t *spgd)
{
/* If the entry's not present, there's nothing to release. */
if (pgd_flags(*spgd) & _PAGE_PRESENT) {
@@ -342,15 +496,18 @@ static void release_pgd(struct lguest *lg, pgd_t *spgd)
}
}
+#endif
+
/*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 < 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,
@@ -370,15 +527,30 @@ 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
+
+
/* First step: get the top-level Guest page table entry. */
- gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
+ gpgd = lgread(cpu, (unsigned long) gpgd_addr(cpu, vaddr), pgd_t);
/* Toplevel not present? We can't map it in. */
if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) {
kill_guest(cpu, "Bad address %#lx", vaddr);
return -1UL;
}
- gpte = lgread(cpu, gpte_addr(gpgd, vaddr), pte_t);
+ gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t);
+
+#ifdef CONFIG_X86_PAE
+ gpmd = lgread(cpu, (unsigned long) gpmd_addr(gpgd, vaddr), pmd_t);
+ if (!(pmd_flags(gpmd) & _PAGE_PRESENT))
+ kill_guest(cpu, "Bad address %#lx", vaddr);
+#endif
+
+ gpte = lgread(cpu, (unsigned long) gpte_addr(cpu, gpgd, vaddr), pte_t);
+
+
if (!(pte_flags(gpte) & _PAGE_PRESENT))
kill_guest(cpu, "Bad address %#lx", vaddr);
@@ -388,7 +560,7 @@ unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr)
/* 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. */
-static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable)
+static unsigned int find_pgdir(struct lguest *lg, pgd_t *pgtable)
{
unsigned int i;
for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
@@ -401,10 +573,13 @@ 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 lg_cpu *cpu,
- unsigned long gpgdir,
+ pgd_t *gpgdir,
int *blank_pgdir)
{
unsigned int next;
+#ifdef CONFIG_X86_PAE
+ pmd_t *pmd_table;
+#endif
/* We pick one entry at random to throw out. Choosing the Least
* Recently Used might be better, but this is easy. */
@@ -413,13 +588,36 @@ static unsigned int new_pgdir(struct lg_cpu *cpu,
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 (!cpu->lg->pgdirs[next].pgdir)
next = cpu->cpu_pgd;
+
+#ifdef CONFIG_X86_PAE
+ else {
+ /* In PAE mode, allocate a pmd page and populate the
+ * last pgd entry. */
+ pmd_table = (pmd_t *) get_zeroed_page(GFP_KERNEL);
+ if (!pmd_table){
+ free_page ((long) cpu->lg->pgdirs[next].pgdir);
+ native_set_pud((pud_t *)cpu->lg->pgdirs[next].pgdir, __pud(0));
+ next = cpu->cpu_pgd;
+ }
+ else {
+
+ native_set_pud((pud_t *) cpu->lg->pgdirs[next].pgdir +
+ SWITCHER_PGD_INDEX,
+ __pud(__pa(pmd_table) | _PAGE_PRESENT));
+
+ /* This is a blank page, so there are no kernel
+ * mappings: caller must map the stack! */
+ *blank_pgdir = 1;
+ }
+ }
+#else
else
- /* This is a blank page, so there are no kernel
- * mappings: caller must map the stack! */
*blank_pgdir = 1;
+#endif
}
/* Record which Guest toplevel this shadows. */
cpu->lg->pgdirs[next].gpgdir = gpgdir;
@@ -431,7 +629,7 @@ static unsigned int new_pgdir(struct lg_cpu *cpu,
/*H:430 (iv) Switching page tables
*
- * Now we've seen all the page table setting and manipulation, let's see what
+ * 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)
@@ -439,11 +637,11 @@ 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(cpu->lg, pgtable);
+ newpgdir = find_pgdir(cpu->lg, (pgd_t *)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);
+ newpgdir = new_pgdir(cpu, (pgd_t *)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 */
@@ -456,14 +654,30 @@ void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable)
* when we destroy the Guest. */
static void release_all_pagetables(struct lguest *lg)
{
- unsigned int i, j;
+ unsigned int i, j, k;
+
+#ifdef CONFIG_X86_PAE
+ pgd_t *spgd;
+ pmd_t *pmdpage;
+#endif
/* Every shadow pagetable this Guest has */
for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
- if (lg->pgdirs[i].pgdir)
+ 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);
+ release_pgd(lg->pgdirs[i].pgdir + j);
+#ifdef CONFIG_X86_PAE
+ /* Get the last pmd page. */
+ spgd = lg->pgdirs[i].pgdir + SWITCHER_PGD_INDEX;
+ pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
+
+ /* And release the pmd entries of that pmd page,
+ * except for the switcher pmd. */
+ for (k = 0; k < SWITCHER_PMD_INDEX; k++)
+ release_pmd(&pmdpage[k]);
+#endif
+ }
}
/* We also throw away everything when a Guest tells us it's changed a kernel
@@ -505,23 +719,39 @@ static void do_set_pte(struct lg_cpu *cpu, int idx,
/* 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, we 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)) {
+ check_gpte(cpu, gpte);
+ native_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. */
+ native_set_pte (spte, __pte(0));
+#ifdef CONFIG_X86_PAE
+ }
+#endif
}
}
@@ -547,7 +777,7 @@ void guest_set_pte(struct lg_cpu *cpu,
do_set_pte(cpu, i, vaddr, gpte);
} else {
/* Is this page table one we have a shadow for? */
- int pgdir = find_pgdir(cpu->lg, gpgdir);
+ int pgdir = find_pgdir(cpu->lg, (pgd_t *)gpgdir);
if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs))
/* If so, do the update. */
do_set_pte(cpu, pgdir, vaddr, gpte);
@@ -568,9 +798,38 @@ void guest_set_pte(struct lg_cpu *cpu,
*
* 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 gpgdir, u32 idx)
+
+#ifdef CONFIG_X86_PAE
+void guest_set_pud(struct lguest *lg, unsigned long pudp, u32 idx)
+{
+ int pgdir;
+ pgd_t *gpgdir = (pgd_t *) pudp;
+
+ if (idx >= SWITCHER_PGD_INDEX){
+ printk ("tryied to map on the last pgd entry\n");
+ 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))
+ /* ... throw it away. */
+ release_pgd(lg->pgdirs[pgdir].pgdir + idx);
+
+}
+
+void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx)
+{
+ guest_pagetable_clear_all(&lg->cpus[0]); //ugly
+}
+
+#else /*!CONFIG_X86_PAE*/
+
+void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx)
{
int pgdir;
+ pgd_t *gpgdir = (pgd_t *) pmdp;
/* The kernel seems to try to initialize this early on: we ignore its
* attempts to map over the Switcher. */
@@ -581,8 +840,9 @@ void guest_set_pmd(struct lguest *lg, unsigned long gpgdir, u32 idx)
pgdir = find_pgdir(lg, gpgdir);
if (pgdir < ARRAY_SIZE(lg->pgdirs))
/* ... throw it away. */
- release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx);
+ release_pgd(lg->pgdirs[pgdir].pgdir + idx);
}
+#endif
/* Once we know how much memory we have we can construct simple identity
* (which set virtual == physical) and linear mappings
@@ -596,8 +856,14 @@ static unsigned long setup_pagetables(struct lguest *lg,
{
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;
+ unsigned int mapped_pages, i, linear_pages;
+#ifdef CONFIG_X86_PAE
+ u64 *pmds;
+ unsigned int j;
+#else
+ unsigned int phys_linear;
+#endif
/* We have mapped_pages frames to map, so we need
* linear_pages page tables to map them. */
@@ -609,6 +875,9 @@ static unsigned long setup_pagetables(struct lguest *lg,
/* Now we use the next linear_pages pages as pte pages */
linear = (void *)pgdir - linear_pages * PAGE_SIZE;
+#ifdef CONFIG_X86_PAE
+ pmds = (void *)linear - PAGE_SIZE;
+#endif
/* Linear mapping is easy: put every page's address into the
* mapping in order. */
@@ -619,8 +888,26 @@ static unsigned long setup_pagetables(struct lguest *lg,
return -EFAULT;
}
+#ifdef CONFIG_X86_PAE
/* The top level points to the linear page table pages above.
* We setup the identity and linear mappings here. */
+ for (i = 0, j = 0; i < mapped_pages; i += PTRS_PER_PTE, j++) {
+ pmd_t pmd;
+ pmd = __pmd( ((unsigned long)(linear+i) - mem_base) |
+ _PAGE_PRESENT | _PAGE_RW | _PAGE_USER);
+ if (copy_to_user(&pmds[j], &pmd, sizeof(pmd)) != 0)
+ return -EFAULT;
+ }
+ pgd_t pgd;
+ pgd = __pgd((((u32)pmds) - mem_base) | _PAGE_PRESENT);
+
+ if (copy_to_user(&pgdir[0], &pgd, sizeof(pgd)) != 0)
+ return -EFAULT;
+
+ if (copy_to_user(&pgdir[3], &pgd, sizeof(pgd)) != 0)
+ return -EFAULT;
+
+#else
phys_linear = (unsigned long)linear - mem_base;
for (i = 0; i < mapped_pages; i += PTRS_PER_PTE) {
pgd_t pgd;
@@ -633,6 +920,7 @@ static unsigned long setup_pagetables(struct lguest *lg,
&pgd, sizeof(pgd)))
return -EFAULT;
}
+#endif
/* We return the top level (guest-physical) address: remember where
* this is. */
@@ -648,6 +936,10 @@ int init_guest_pagetable(struct lguest *lg)
u64 mem;
u32 initrd_size;
struct boot_params __user *boot = (struct boot_params *)lg->mem_base;
+#ifdef CONFIG_X86_PAE
+ pgd_t *pgd;
+ pmd_t *pmd_table;
+#endif
/* Get the Guest memory size and the ramdisk size from the boot header
* located at lg->mem_base (Guest address 0). */
@@ -657,12 +949,23 @@ int init_guest_pagetable(struct lguest *lg)
/* We start on the first shadow page table, and give it a blank PGD
* page. */
- 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].gpgdir = (pgd_t *) setup_pagetables(lg, mem, initrd_size);
+ if (IS_ERR_VALUE((int) lg->pgdirs[0].gpgdir))
+ return (int) lg->pgdirs[0].gpgdir;
lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL);
if (!lg->pgdirs[0].pgdir)
return -ENOMEM;
+#ifdef CONFIG_X86_PAE
+ pgd = lg->pgdirs[0].pgdir;
+ pmd_table = (pmd_t *) get_zeroed_page(GFP_KERNEL);
+ if (!pmd_table)
+ return -ENOMEM;
+
+ native_set_pud((pud_t *) pgd + SWITCHER_PGD_INDEX,
+ __pud(__pa(pmd_table) | _PAGE_PRESENT));
+
+#endif
+
lg->cpus[0].cpu_pgd = 0;
return 0;
}
@@ -670,21 +973,36 @@ int init_guest_pagetable(struct lguest *lg)
/* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
void page_table_guest_data_init(struct lg_cpu *cpu)
{
+#ifdef CONFIG_X86_PAE
+ const unsigned long reserve_mb = 2;
+#else
+ const unsigned long reserve_mb = 4;
+#endif
+
/* 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))
+ &cpu->lg->lguest_data->kernel_address)
+ /* We tell the Guest that it can't use the top 2 or 4 MB
+ * of virtual addresses used by the Switcher. */
+ || put_user(reserve_mb * 1024 * 1024,
+ &cpu->lg->lguest_data->reserve_mem)
+ || put_user((unsigned long) 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. */
+#ifdef CONFIG_X86_PAE
+ if (pgd_index(cpu->lg->kernel_address) == SWITCHER_PGD_INDEX)
+ if (pmd_index(cpu->lg->kernel_address) == SWITCHER_PMD_INDEX)
+ kill_guest(cpu, "bad kernel address %#lx",
+ cpu->lg->kernel_address);
+#else
if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX)
kill_guest(cpu, "bad kernel address %#lx",
cpu->lg->kernel_address);
+#endif
}
/* When a Guest dies, our cleanup is fairly simple. */
@@ -708,15 +1026,28 @@ void free_guest_pagetable(struct lguest *lg)
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;
+#ifdef CONFIG_X86_PAE
+ pmd_t switcher_pmd;
+ pmd_t *pmd_table;
+
+ switcher_pmd = pfn_pmd(__pa(switcher_pte_page) >>
+ PAGE_SHIFT, __pgprot(__PAGE_KERNEL));
+ pmd_table = __va(pgd_pfn(cpu->lg->
+ pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX])
+ << PAGE_SHIFT);
+ pmd_table[SWITCHER_PMD_INDEX] = switcher_pmd;
+
+#else
+ pgd_t switcher_pgd;
+
/* 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;
+#endif
/* 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
@@ -727,7 +1058,8 @@ void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages)
* 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;
+ switcher_pte_page[(unsigned long)pages / PAGE_SIZE % PTRS_PER_PTE]
+ = regs_pte;
}
/*:*/
@@ -752,21 +1084,23 @@ static __init void populate_switcher_pte_page(unsigned int cpu,
/* 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));
+ native_set_pte(&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_pte(page_to_pfn(switcher_page[i]),
- __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW));
+ native_set_pte(&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));
+ native_set_pte(&pte[i+1],pfn_pte(page_to_pfn(switcher_page[i+1]),
+ __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)) );
+
+// look rules for set_pte at pgtable-3level.h
}
/* We've made it through the page table code. Perhaps our tired brains are
More information about the Lguest
mailing list