Merge git://git.kernel.org/pub/scm/linux/kernel/git/rusty/linux-2.6-lguest

* git://git.kernel.org/pub/scm/linux/kernel/git/rusty/linux-2.6-lguest: (31 commits)
  lguest: add support for indirect ring entries
  lguest: suppress notifications in example Launcher
  lguest: try to batch interrupts on network receive
  lguest: avoid sending interrupts to Guest when no activity occurs.
  lguest: implement deferred interrupts in example Launcher
  lguest: remove obsolete LHREQ_BREAK call
  lguest: have example Launcher service all devices in separate threads
  lguest: use eventfds for device notification
  eventfd: export eventfd_signal and eventfd_fget for lguest
  lguest: allow any process to send interrupts
  lguest: PAE fixes
  lguest: PAE support
  lguest: Add support for kvm_hypercall4()
  lguest: replace hypercall name LHCALL_SET_PMD with LHCALL_SET_PGD
  lguest: use native_set_* macros, which properly handle 64-bit entries when PAE is activated
  lguest: map switcher with executable page table entries
  lguest: fix writev returning short on console output
  lguest: clean up length-used value in example launcher
  lguest: Segment selectors are 16-bit long. Fix lg_cpu.ss1 definition.
  lguest: beyond ARRAY_SIZE of cpu->arch.gdt
  ...

Documentation/lguest/Makefile

@@ -1,6 +1,5 @@
 # This creates the demonstration utility "lguest" which runs a Linux guest.
-CFLAGS:=-Wall -Wmissing-declarations -Wmissing-prototypes -O3 -I../../include -I../../arch/x86/include -U_FORTIFY_SOURCE
+CFLAGS:=-m32 -Wall -Wmissing-declarations -Wmissing-prototypes -O3 -I../../include -I../../arch/x86/include -U_FORTIFY_SOURCE
-LDLIBS:=-lz
 
 all: lguest
 

Documentation/lguest/lguest.txt

@@ -37,7 +37,6 @@ Running Lguest:
      "Paravirtualized guest support" = Y
         "Lguest guest support" = Y
      "High Memory Support" = off/4GB
-     "PAE (Physical Address Extension) Support" = N
      "Alignment value to which kernel should be aligned" = 0x100000
         (CONFIG_PARAVIRT=y, CONFIG_LGUEST_GUEST=y, CONFIG_HIGHMEM64G=n and
          CONFIG_PHYSICAL_ALIGN=0x100000)

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)
 
-/* We map at -4M for ease of mapping into the guest (one PTE page). */
+/* We map at -4M (-2M when PAE is activated) for ease of mapping
+ * into the guest (one PTE page). */
+#ifdef CONFIG_X86_PAE
+#define SWITCHER_ADDR 0xFFE00000
+#else
 #define SWITCHER_ADDR 0xFFC00000
+#endif
 
 /* Found in switcher.S */
 extern unsigned long default_idt_entries[];

arch/x86/include/asm/lguest_hcall.h

@@ -12,11 +12,13 @@
 #define LHCALL_TS		8
 #define LHCALL_SET_CLOCKEVENT	9
 #define LHCALL_HALT		10
+#define LHCALL_SET_PMD		13
 #define LHCALL_SET_PTE		14
-#define LHCALL_SET_PMD		15
+#define LHCALL_SET_PGD		15
 #define LHCALL_LOAD_TLS		16
 #define LHCALL_NOTIFY		17
 #define LHCALL_LOAD_GDT_ENTRY	18
+#define LHCALL_SEND_INTERRUPTS	19
 
 #define LGUEST_TRAP_ENTRY 0x1F
 
@@ -32,10 +34,10 @@
  * operations?  There are two ways: the direct way is to make a "hypercall",
  * to make requests of the Host Itself.
  *
- * We use the KVM hypercall mechanism. Eighteen hypercalls are
+ * We use the KVM hypercall mechanism. Seventeen hypercalls are
  * available: the hypercall number is put in the %eax register, and the
- * arguments (when required) are placed in %ebx, %ecx and %edx.  If a return
+ * arguments (when required) are placed in %ebx, %ecx, %edx and %esi.
- * value makes sense, it's returned in %eax.
+ * If a return value makes sense, it's returned in %eax.
  *
  * Grossly invalid calls result in Sudden Death at the hands of the vengeful
  * Host, rather than returning failure.  This reflects Winston Churchill's
@@ -47,8 +49,9 @@
 
 #define LHCALL_RING_SIZE 64
 struct hcall_args {
-	/* These map directly onto eax, ebx, ecx, edx in struct lguest_regs */
+	/* These map directly onto eax, ebx, ecx, edx and esi
-	unsigned long arg0, arg1, arg2, arg3;
+	 * in struct lguest_regs */
+	unsigned long arg0, arg1, arg2, arg3, arg4;
 };
 
 #endif /* !__ASSEMBLY__ */

arch/x86/kernel/asm-offsets_32.c

@@ -126,6 +126,7 @@ void foo(void)
 #if defined(CONFIG_LGUEST) || defined(CONFIG_LGUEST_GUEST) || defined(CONFIG_LGUEST_MODULE)
 	BLANK();
 	OFFSET(LGUEST_DATA_irq_enabled, lguest_data, irq_enabled);
+	OFFSET(LGUEST_DATA_irq_pending, lguest_data, irq_pending);
 	OFFSET(LGUEST_DATA_pgdir, lguest_data, pgdir);
 
 	BLANK();

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

arch/x86/lguest/boot.c

@@ -87,7 +87,7 @@ struct lguest_data lguest_data = {
 
 /*G:037 async_hcall() is pretty simple: I'm quite proud of it really.  We have a
  * ring buffer of stored hypercalls which the Host will run though next time we
- * do a normal hypercall.  Each entry in the ring has 4 slots for the hypercall
+ * do a normal hypercall.  Each entry in the ring has 5 slots for the hypercall
  * arguments, and a "hcall_status" word which is 0 if the call is ready to go,
  * and 255 once the Host has finished with it.
  *
@@ -96,7 +96,8 @@ struct lguest_data lguest_data = {
  * effect of causing the Host to run all the stored calls in the ring buffer
  * which empties it for next time! */
 static void async_hcall(unsigned long call, unsigned long arg1,
-			unsigned long arg2, unsigned long arg3)
+			unsigned long arg2, unsigned long arg3,
+			unsigned long arg4)
 {
 	/* Note: This code assumes we're uniprocessor. */
 	static unsigned int next_call;
@@ -108,12 +109,13 @@ static void async_hcall(unsigned long call, unsigned long arg1,
 	local_irq_save(flags);
 	if (lguest_data.hcall_status[next_call] != 0xFF) {
 		/* Table full, so do normal hcall which will flush table. */
-		kvm_hypercall3(call, arg1, arg2, arg3);
+		kvm_hypercall4(call, arg1, arg2, arg3, arg4);
 	} else {
 		lguest_data.hcalls[next_call].arg0 = call;
 		lguest_data.hcalls[next_call].arg1 = arg1;
 		lguest_data.hcalls[next_call].arg2 = arg2;
 		lguest_data.hcalls[next_call].arg3 = arg3;
+		lguest_data.hcalls[next_call].arg4 = arg4;
 		/* Arguments must all be written before we mark it to go */
 		wmb();
 		lguest_data.hcall_status[next_call] = 0;
@@ -141,7 +143,7 @@ static void lazy_hcall1(unsigned long call,
 	if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
 		kvm_hypercall1(call, arg1);
 	else
-		async_hcall(call, arg1, 0, 0);
+		async_hcall(call, arg1, 0, 0, 0);
 }
 
 static void lazy_hcall2(unsigned long call,
@@ -151,7 +153,7 @@ static void lazy_hcall2(unsigned long call,
 	if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
 		kvm_hypercall2(call, arg1, arg2);
 	else
-		async_hcall(call, arg1, arg2, 0);
+		async_hcall(call, arg1, arg2, 0, 0);
 }
 
 static void lazy_hcall3(unsigned long call,
@@ -162,9 +164,23 @@ static void lazy_hcall3(unsigned long call,
 	if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
 		kvm_hypercall3(call, arg1, arg2, arg3);
 	else
-		async_hcall(call, arg1, arg2, arg3);
+		async_hcall(call, arg1, arg2, arg3, 0);
 }
 
+#ifdef CONFIG_X86_PAE
+static void lazy_hcall4(unsigned long call,
+		       unsigned long arg1,
+		       unsigned long arg2,
+		       unsigned long arg3,
+		       unsigned long arg4)
+{
+	if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_NONE)
+		kvm_hypercall4(call, arg1, arg2, arg3, arg4);
+	else
+		async_hcall(call, arg1, arg2, arg3, arg4);
+}
+#endif
+
 /* When lazy mode is turned off reset the per-cpu lazy mode variable and then
  * issue the do-nothing hypercall to flush any stored calls. */
 static void lguest_leave_lazy_mmu_mode(void)
@@ -179,7 +195,7 @@ static void lguest_end_context_switch(struct task_struct *next)
 	paravirt_end_context_switch(next);
 }
 
-/*G:033
+/*G:032
  * After that diversion we return to our first native-instruction
  * replacements: four functions for interrupt control.
  *
@@ -199,30 +215,28 @@ static unsigned long save_fl(void)
 {
 	return lguest_data.irq_enabled;
 }
-PV_CALLEE_SAVE_REGS_THUNK(save_fl);
-
-/* restore_flags() just sets the flags back to the value given. */
-static void restore_fl(unsigned long flags)
-{
-	lguest_data.irq_enabled = flags;
-}
-PV_CALLEE_SAVE_REGS_THUNK(restore_fl);
 
 /* Interrupts go off... */
 static void irq_disable(void)
 {
 	lguest_data.irq_enabled = 0;
 }
+
+/* Let's pause a moment.  Remember how I said these are called so often?
+ * Jeremy Fitzhardinge optimized them so hard early in 2009 that he had to
+ * break some rules.  In particular, these functions are assumed to save their
+ * own registers if they need to: normal C functions assume they can trash the
+ * eax register.  To use normal C functions, we use
+ * PV_CALLEE_SAVE_REGS_THUNK(), which pushes %eax onto the stack, calls the
+ * C function, then restores it. */
+PV_CALLEE_SAVE_REGS_THUNK(save_fl);
 PV_CALLEE_SAVE_REGS_THUNK(irq_disable);
-
-/* Interrupts go on... */
-static void irq_enable(void)
-{
-	lguest_data.irq_enabled = X86_EFLAGS_IF;
-}
-PV_CALLEE_SAVE_REGS_THUNK(irq_enable);
-
 /*:*/
+
+/* These are in i386_head.S */
+extern void lg_irq_enable(void);
+extern void lg_restore_fl(unsigned long flags);
+
 /*M:003 Note that we don't check for outstanding interrupts when we re-enable
  * them (or when we unmask an interrupt).  This seems to work for the moment,
  * since interrupts are rare and we'll just get the interrupt on the next timer
@@ -368,8 +382,8 @@ 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. */
-		*dx &= 0x07808111;
+		*dx &= 0x07808151;
 		/* 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
@@ -388,6 +402,11 @@ static void lguest_cpuid(unsigned int *ax, unsigned int *bx,
 		if (*ax > 0x80000008)
 			*ax = 0x80000008;
 		break;
+	case 0x80000001:
+		/* Here we should fix nx cap depending on host. */
+		/* For this version of PAE, we just clear NX bit. */
+		*dx &= ~(1 << 20);
+		break;
 	}
 }
 
@@ -521,25 +540,52 @@ static void lguest_write_cr4(unsigned long val)
 static void lguest_pte_update(struct mm_struct *mm, unsigned long addr,
 			       pte_t *ptep)
 {
+#ifdef CONFIG_X86_PAE
+	lazy_hcall4(LHCALL_SET_PTE, __pa(mm->pgd), addr,
+		    ptep->pte_low, ptep->pte_high);
+#else
 	lazy_hcall3(LHCALL_SET_PTE, __pa(mm->pgd), addr, ptep->pte_low);
+#endif
 }
 
 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);
 }
 
-/* The Guest calls this to set a top-level entry.  Again, we set the entry then
+/* The Guest calls lguest_set_pud to set a top-level entry and lguest_set_pmd
- * tell the Host which top-level page we changed, and the index of the entry we
+ * to set a middle-level entry when PAE is activated.
- * changed. */
+ * Again, we set the entry then tell the Host which page we changed,
+ * and the index of the entry we changed. */
+#ifdef CONFIG_X86_PAE
+static void lguest_set_pud(pud_t *pudp, pud_t pudval)
+{
+	native_set_pud(pudp, pudval);
+
+	/* 32 bytes aligned pdpt address and the index. */
+	lazy_hcall2(LHCALL_SET_PGD, __pa(pudp) & 0xFFFFFFE0,
+		   (__pa(pudp) & 0x1F) / sizeof(pud_t));
+}
+
 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));
 }
+#else
+
+/* The Guest calls lguest_set_pmd to set a top-level entry when PAE is not
+ * activated. */
+static void lguest_set_pmd(pmd_t *pmdp, pmd_t pmdval)
+{
+	native_set_pmd(pmdp, pmdval);
+	lazy_hcall2(LHCALL_SET_PGD, __pa(pmdp) & PAGE_MASK,
+		   (__pa(pmdp) & (PAGE_SIZE - 1)) / sizeof(pmd_t));
+}
+#endif
 
 /* There are a couple of legacy places where the kernel sets a PTE, but we
  * don't know the top level any more.  This is useless for us, since we don't
@@ -552,11 +598,31 @@ 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);
+	lguest_pte_update(mm, addr, ptep);
+}
+
+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
@@ -628,13 +694,12 @@ static void __init lguest_init_IRQ(void)
 {
 	unsigned int i;
 
-	for (i = 0; i < LGUEST_IRQS; i++) {
+	for (i = FIRST_EXTERNAL_VECTOR; i < NR_VECTORS; i++) {
-		int vector = FIRST_EXTERNAL_VECTOR + i;
 		/* Some systems map "vectors" to interrupts weirdly.  Lguest has
 		 * a straightforward 1 to 1 mapping, so force that here. */
-		__get_cpu_var(vector_irq)[vector] = i;
+		__get_cpu_var(vector_irq)[i] = i - FIRST_EXTERNAL_VECTOR;
-		if (vector != SYSCALL_VECTOR)
+		if (i != SYSCALL_VECTOR)
-			set_intr_gate(vector, interrupt[i]);
+			set_intr_gate(i, interrupt[i - FIRST_EXTERNAL_VECTOR]);
 	}
 	/* This call is required to set up for 4k stacks, where we have
 	 * separate stacks for hard and soft interrupts. */
@@ -973,10 +1038,10 @@ static void lguest_restart(char *reason)
  *
  * Our current solution is to allow the paravirt back end to optionally patch
  * over the indirect calls to replace them with something more efficient.  We
- * patch the four most commonly called functions: disable interrupts, enable
+ * patch two of the simplest of the most commonly called functions: disable
- * interrupts, restore interrupts and save interrupts.  We usually have 6 or 10
+ * interrupts and save interrupts.  We usually have 6 or 10 bytes to patch
- * bytes to patch into: the Guest versions of these operations are small enough
+ * into: the Guest versions of these operations are small enough that we can
- * that we can fit comfortably.
+ * fit comfortably.
  *
  * First we need assembly templates of each of the patchable Guest operations,
  * and these are in i386_head.S. */
@@ -987,8 +1052,6 @@ static const struct lguest_insns
 	const char *start, *end;
 } lguest_insns[] = {
 	[PARAVIRT_PATCH(pv_irq_ops.irq_disable)] = { lgstart_cli, lgend_cli },
-	[PARAVIRT_PATCH(pv_irq_ops.irq_enable)] = { lgstart_sti, lgend_sti },
-	[PARAVIRT_PATCH(pv_irq_ops.restore_fl)] = { lgstart_popf, lgend_popf },
 	[PARAVIRT_PATCH(pv_irq_ops.save_fl)] = { lgstart_pushf, lgend_pushf },
 };
 
@@ -1026,6 +1089,7 @@ __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. */
@@ -1033,9 +1097,9 @@ __init void lguest_init(void)
 	/* interrupt-related operations */
 	pv_irq_ops.init_IRQ = lguest_init_IRQ;
 	pv_irq_ops.save_fl = PV_CALLEE_SAVE(save_fl);
-	pv_irq_ops.restore_fl = PV_CALLEE_SAVE(restore_fl);
+	pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(lg_restore_fl);
 	pv_irq_ops.irq_disable = PV_CALLEE_SAVE(irq_disable);
-	pv_irq_ops.irq_enable = PV_CALLEE_SAVE(irq_enable);
+	pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(lg_irq_enable);
 	pv_irq_ops.safe_halt = lguest_safe_halt;
 
 	/* init-time operations */
@@ -1071,6 +1135,12 @@ __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;

arch/x86/lguest/i386_head.S

@@ -46,10 +46,64 @@ ENTRY(lguest_entry)
 	.globl lgstart_##name; .globl lgend_##name
 
 LGUEST_PATCH(cli, movl $0, lguest_data+LGUEST_DATA_irq_enabled)
-LGUEST_PATCH(sti, movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled)
-LGUEST_PATCH(popf, movl %eax, lguest_data+LGUEST_DATA_irq_enabled)
 LGUEST_PATCH(pushf, movl lguest_data+LGUEST_DATA_irq_enabled, %eax)
-/*:*/
+
+/*G:033 But using those wrappers is inefficient (we'll see why that doesn't
+ * matter for save_fl and irq_disable later).  If we write our routines
+ * carefully in assembler, we can avoid clobbering any registers and avoid
+ * jumping through the wrapper functions.
+ *
+ * I skipped over our first piece of assembler, but this one is worth studying
+ * in a bit more detail so I'll describe in easy stages.  First, the routine
+ * to enable interrupts: */
+ENTRY(lg_irq_enable)
+	/* The reverse of irq_disable, this sets lguest_data.irq_enabled to
+	 * X86_EFLAGS_IF (ie. "Interrupts enabled"). */
+	movl $X86_EFLAGS_IF, lguest_data+LGUEST_DATA_irq_enabled
+	/* But now we need to check if the Host wants to know: there might have
+	 * been interrupts waiting to be delivered, in which case it will have
+	 * set lguest_data.irq_pending to X86_EFLAGS_IF.  If it's not zero, we
+	 * jump to send_interrupts, otherwise we're done. */
+	testl $0, lguest_data+LGUEST_DATA_irq_pending
+	jnz send_interrupts
+	/* One cool thing about x86 is that you can do many things without using
+	 * a register.  In this case, the normal path hasn't needed to save or
+	 * restore any registers at all! */
+	ret
+send_interrupts:
+	/* OK, now we need a register: eax is used for the hypercall number,
+	 * which is LHCALL_SEND_INTERRUPTS.
+	 *
+	 * We used not to bother with this pending detection at all, which was
+	 * much simpler.  Sooner or later the Host would realize it had to
+	 * send us an interrupt.  But that turns out to make performance 7
+	 * times worse on a simple tcp benchmark.  So now we do this the hard
+	 * way. */
+	pushl %eax
+	movl $LHCALL_SEND_INTERRUPTS, %eax
+	/* This is a vmcall instruction (same thing that KVM uses).  Older
+	 * assembler versions might not know the "vmcall" instruction, so we
+	 * create one manually here. */
+	.byte 0x0f,0x01,0xc1 /* KVM_HYPERCALL */
+	popl %eax
+	ret
+
+/* Finally, the "popf" or "restore flags" routine.  The %eax register holds the
+ * flags (in practice, either X86_EFLAGS_IF or 0): if it's X86_EFLAGS_IF we're
+ * enabling interrupts again, if it's 0 we're leaving them off. */
+ENTRY(lg_restore_fl)
+	/* This is just "lguest_data.irq_enabled = flags;" */
+	movl %eax, lguest_data+LGUEST_DATA_irq_enabled
+	/* Now, if the %eax value has enabled interrupts and
+	 * lguest_data.irq_pending is set, we want to tell the Host so it can
+	 * deliver any outstanding interrupts.  Fortunately, both values will
+	 * be X86_EFLAGS_IF (ie. 512) in that case, and the "testl"
+	 * instruction will AND them together for us.  If both are set, we
+	 * jump to send_interrupts. */
+	testl lguest_data+LGUEST_DATA_irq_pending, %eax
+	jnz send_interrupts
+	/* Again, the normal path has used no extra registers.  Clever, huh? */
+	ret
 
 /* These demark the EIP range where host should never deliver interrupts. */
 .global lguest_noirq_start

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 && EVENTFD
 	select HVC_DRIVER
 	---help---
 	  This is a very simple module which allows you to run

drivers/lguest/core.c

@@ -95,7 +95,7 @@ static __init int map_switcher(void)
 	 * array of struct pages.  It increments that pointer, but we don't
 	 * care. */
 	pagep = switcher_page;
-	err = map_vm_area(switcher_vma, PAGE_KERNEL, &pagep);
+	err = map_vm_area(switcher_vma, PAGE_KERNEL_EXEC, &pagep);
 	if (err) {
 		printk("lguest: map_vm_area failed: %i\n", err);
 		goto free_vma;
@@ -188,6 +188,9 @@ int run_guest(struct lg_cpu *cpu, unsigned long __user *user)
 {
 	/* We stop running once the Guest is dead. */
 	while (!cpu->lg->dead) {
+		unsigned int irq;
+		bool more;
+
 		/* First we run any hypercalls the Guest wants done. */
 		if (cpu->hcall)
 			do_hypercalls(cpu);
@@ -195,23 +198,23 @@ int run_guest(struct lg_cpu *cpu, unsigned long __user *user)
 		/* It's possible the Guest did a NOTIFY hypercall to the
 		 * Launcher, in which case we return from the read() now. */
 		if (cpu->pending_notify) {
+			if (!send_notify_to_eventfd(cpu)) {
 				if (put_user(cpu->pending_notify, user))
 					return -EFAULT;
 				return sizeof(cpu->pending_notify);
 			}
+		}
 
 		/* Check for signals */
 		if (signal_pending(current))
 			return -ERESTARTSYS;
 
-		/* If Waker set break_out, return to Launcher. */
-		if (cpu->break_out)
-			return -EAGAIN;
-
 		/* Check if there are any interrupts which can be delivered now:
 		 * if so, this sets up the hander to be executed when we next
 		 * run the Guest. */
-		maybe_do_interrupt(cpu);
+		irq = interrupt_pending(cpu, &more);
+		if (irq < LGUEST_IRQS)
+			try_deliver_interrupt(cpu, irq, more);
 
 		/* All long-lived kernel loops need to check with this horrible
 		 * thing called the freezer.  If the Host is trying to suspend,
@@ -224,9 +227,14 @@ int run_guest(struct lg_cpu *cpu, unsigned long __user *user)
 			break;
 
 		/* If the Guest asked to be stopped, we sleep.  The Guest's
-		 * clock timer or LHREQ_BREAK from the Waker will wake us. */
+		 * clock timer will wake us. */
 		if (cpu->halted) {
 			set_current_state(TASK_INTERRUPTIBLE);
+			/* Just before we sleep, make sure no interrupt snuck in
+			 * which we should be doing. */
+			if (interrupt_pending(cpu, &more) < LGUEST_IRQS)
+				set_current_state(TASK_RUNNING);
+			else
 				schedule();
 			continue;
 		}

drivers/lguest/hypercalls.c

@@ -37,6 +37,10 @@ static void do_hcall(struct lg_cpu *cpu, struct hcall_args *args)
 		/* This call does nothing, except by breaking out of the Guest
 		 * it makes us process all the asynchronous hypercalls. */
 		break;
+	case LHCALL_SEND_INTERRUPTS:
+		/* This call does nothing too, but by breaking out of the Guest
+		 * it makes us process any pending interrupts. */
+		break;
 	case LHCALL_LGUEST_INIT:
 		/* You can't get here unless you're already initialized.  Don't
 		 * do that. */
@@ -73,11 +77,21 @@ static void do_hcall(struct lg_cpu *cpu, struct hcall_args *args)
 		guest_set_stack(cpu, args->arg1, args->arg2, args->arg3);
 		break;
 	case LHCALL_SET_PTE:
+#ifdef CONFIG_X86_PAE
+		guest_set_pte(cpu, args->arg1, args->arg2,
+				__pte(args->arg3 | (u64)args->arg4 << 32));
+#else
 		guest_set_pte(cpu, args->arg1, args->arg2, __pte(args->arg3));
+#endif
 		break;
+	case LHCALL_SET_PGD:
+		guest_set_pgd(cpu->lg, args->arg1, args->arg2);
+		break;
+#ifdef CONFIG_X86_PAE
 	case LHCALL_SET_PMD:
 		guest_set_pmd(cpu->lg, args->arg1, args->arg2);
 		break;
+#endif
 	case LHCALL_SET_CLOCKEVENT:
 		guest_set_clockevent(cpu, args->arg1);
 		break;

drivers/lguest/interrupts_and_traps.c

@@ -128,30 +128,39 @@ static void set_guest_interrupt(struct lg_cpu *cpu, u32 lo, u32 hi,
 /*H:205
  * Virtual Interrupts.
  *
- * maybe_do_interrupt() gets called before every entry to the Guest, to see if
+ * interrupt_pending() returns the first pending interrupt which isn't blocked
- * we should divert the Guest to running an interrupt handler. */
+ * by the Guest.  It is called before every entry to the Guest, and just before
-void maybe_do_interrupt(struct lg_cpu *cpu)
+ * we go to sleep when the Guest has halted itself. */
+unsigned int interrupt_pending(struct lg_cpu *cpu, bool *more)
 {
 	unsigned int irq;
 	DECLARE_BITMAP(blk, LGUEST_IRQS);
-	struct desc_struct *idt;
 
 	/* If the Guest hasn't even initialized yet, we can do nothing. */
 	if (!cpu->lg->lguest_data)
-		return;
+		return LGUEST_IRQS;
 
 	/* Take our "irqs_pending" array and remove any interrupts the Guest
 	 * wants blocked: the result ends up in "blk". */
 	if (copy_from_user(&blk, cpu->lg->lguest_data->blocked_interrupts,
 			   sizeof(blk)))
-		return;
+		return LGUEST_IRQS;
 	bitmap_andnot(blk, cpu->irqs_pending, blk, LGUEST_IRQS);
 
 	/* Find the first interrupt. */
 	irq = find_first_bit(blk, LGUEST_IRQS);
-	/* None?  Nothing to do */
+	*more = find_next_bit(blk, LGUEST_IRQS, irq+1);
-	if (irq >= LGUEST_IRQS)
+
-		return;
+	return irq;
+}
+
+/* This actually diverts the Guest to running an interrupt handler, once an
+ * interrupt has been identified by interrupt_pending(). */
+void try_deliver_interrupt(struct lg_cpu *cpu, unsigned int irq, bool more)
+{
+	struct desc_struct *idt;
+
+	BUG_ON(irq >= LGUEST_IRQS);
 
 	/* They may be in the middle of an iret, where they asked us never to
 	 * deliver interrupts. */
@@ -170,9 +179,13 @@ void maybe_do_interrupt(struct lg_cpu *cpu)
 		u32 irq_enabled;
 		if (get_user(irq_enabled, &cpu->lg->lguest_data->irq_enabled))
 			irq_enabled = 0;
-		if (!irq_enabled)
+		if (!irq_enabled) {
+			/* Make sure they know an IRQ is pending. */
+			put_user(X86_EFLAGS_IF,
+				 &cpu->lg->lguest_data->irq_pending);
 			return;
 		}
+	}
 
 	/* Look at the IDT entry the Guest gave us for this interrupt.  The
 	 * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
@@ -194,6 +207,25 @@ void maybe_do_interrupt(struct lg_cpu *cpu)
 	 * here is a compromise which means at least it gets updated every
 	 * timer interrupt. */
 	write_timestamp(cpu);
+
+	/* If there are no other interrupts we want to deliver, clear
+	 * the pending flag. */
+	if (!more)
+		put_user(0, &cpu->lg->lguest_data->irq_pending);
+}
+
+/* And this is the routine when we want to set an interrupt for the Guest. */
+void set_interrupt(struct lg_cpu *cpu, unsigned int irq)
+{
+	/* Next time the Guest runs, the core code will see if it can deliver
+	 * this interrupt. */
+	set_bit(irq, cpu->irqs_pending);
+
+	/* Make sure it sees it; it might be asleep (eg. halted), or
+	 * running the Guest right now, in which case kick_process()
+	 * will knock it out. */
+	if (!wake_up_process(cpu->tsk))
+		kick_process(cpu->tsk);
 }
 /*:*/
 
@@ -510,10 +542,7 @@ static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
 	struct lg_cpu *cpu = container_of(timer, struct lg_cpu, hrt);
 
 	/* Remember the first interrupt is the timer interrupt. */
-	set_bit(0, cpu->irqs_pending);
+	set_interrupt(cpu, 0);
-	/* If the Guest is actually stopped, we need to wake it up. */
-	if (cpu->halted)
-		wake_up_process(cpu->tsk);
 	return HRTIMER_NORESTART;
 }
 

drivers/lguest/lg.h

@@ -49,7 +49,7 @@ struct lg_cpu {
 	u32 cr2;
 	int ts;
 	u32 esp1;
-	u8 ss1;
+	u16 ss1;
 
 	/* Bitmap of what has changed: see CHANGED_* above. */
 	int changed;
@@ -71,9 +71,7 @@ struct lg_cpu {
 	/* Virtual clock device */
 	struct hrtimer hrt;
 
-	/* Do we need to stop what we're doing and return to userspace? */
+	/* Did the Guest tell us to halt? */
-	int break_out;
-	wait_queue_head_t break_wq;
 	int halted;
 
 	/* Pending virtual interrupts */
@@ -82,6 +80,16 @@ struct lg_cpu {
 	struct lg_cpu_arch arch;
 };
 
+struct lg_eventfd {
+	unsigned long addr;
+	struct file *event;
+};
+
+struct lg_eventfd_map {
+	unsigned int num;
+	struct lg_eventfd map[];
+};
+
 /* The private info the thread maintains about the guest. */
 struct lguest
 {
@@ -102,6 +110,8 @@ struct lguest
 	unsigned int stack_pages;
 	u32 tsc_khz;
 
+	struct lg_eventfd_map *eventfds;
+
 	/* Dead? */
 	const char *dead;
 };
@@ -137,9 +147,13 @@ 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);
+unsigned int interrupt_pending(struct lg_cpu *cpu, bool *more);
+void try_deliver_interrupt(struct lg_cpu *cpu, unsigned int irq, bool more);
+void set_interrupt(struct lg_cpu *cpu, unsigned int irq);
 bool deliver_trap(struct lg_cpu *cpu, unsigned int num);
 void load_guest_idt_entry(struct lg_cpu *cpu, unsigned int i,
 			  u32 low, u32 hi);
@@ -150,6 +164,7 @@ void setup_default_idt_entries(struct lguest_ro_state *state,
 void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt,
 		const unsigned long *def);
 void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta);
+bool send_notify_to_eventfd(struct lg_cpu *cpu);
 void init_clockdev(struct lg_cpu *cpu);
 bool check_syscall_vector(struct lguest *lg);
 int init_interrupts(void);
@@ -168,7 +183,10 @@ void copy_gdt_tls(const struct lg_cpu *cpu, struct desc_struct *gdt);
 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_pgd(struct lguest *lg, unsigned long gpgdir, u32 i);
+#ifdef CONFIG_X86_PAE
 void guest_set_pmd(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,

drivers/lguest/lguest_user.c

@@ -7,32 +7,83 @@
 #include <linux/miscdevice.h>
 #include <linux/fs.h>
 #include <linux/sched.h>
+#include <linux/eventfd.h>
+#include <linux/file.h>
 #include "lg.h"
 
-/*L:055 When something happens, the Waker process needs a way to stop the
+bool send_notify_to_eventfd(struct lg_cpu *cpu)
- * kernel running the Guest and return to the Launcher.  So the Waker writes
- * LHREQ_BREAK and the value "1" to /dev/lguest to do this.  Once the Launcher
- * has done whatever needs attention, it writes LHREQ_BREAK and "0" to release
- * the Waker. */
-static int break_guest_out(struct lg_cpu *cpu, const unsigned long __user*input)
 {
-	unsigned long on;
+	unsigned int i;
+	struct lg_eventfd_map *map;
 
-	/* Fetch whether they're turning break on or off. */
+	/* lg->eventfds is RCU-protected */
-	if (get_user(on, input) != 0)
+	rcu_read_lock();
+	map = rcu_dereference(cpu->lg->eventfds);
+	for (i = 0; i < map->num; i++) {
+		if (map->map[i].addr == cpu->pending_notify) {
+			eventfd_signal(map->map[i].event, 1);
+			cpu->pending_notify = 0;
+			break;
+		}
+	}
+	rcu_read_unlock();
+	return cpu->pending_notify == 0;
+}
+
+static int add_eventfd(struct lguest *lg, unsigned long addr, int fd)
+{
+	struct lg_eventfd_map *new, *old = lg->eventfds;
+
+	if (!addr)
+		return -EINVAL;
+
+	/* Replace the old array with the new one, carefully: others can
+	 * be accessing it at the same time */
+	new = kmalloc(sizeof(*new) + sizeof(new->map[0]) * (old->num + 1),
+		      GFP_KERNEL);
+	if (!new)
+		return -ENOMEM;
+
+	/* First make identical copy. */
+	memcpy(new->map, old->map, sizeof(old->map[0]) * old->num);
+	new->num = old->num;
+
+	/* Now append new entry. */
+	new->map[new->num].addr = addr;
+	new->map[new->num].event = eventfd_fget(fd);
+	if (IS_ERR(new->map[new->num].event)) {
+		kfree(new);
+		return PTR_ERR(new->map[new->num].event);
+	}
+	new->num++;
+
+	/* Now put new one in place. */
+	rcu_assign_pointer(lg->eventfds, new);
+
+	/* We're not in a big hurry.  Wait until noone's looking at old
+	 * version, then delete it. */
+	synchronize_rcu();
+	kfree(old);
+
+	return 0;
+}
+
+static int attach_eventfd(struct lguest *lg, const unsigned long __user *input)
+{
+	unsigned long addr, fd;
+	int err;
+
+	if (get_user(addr, input) != 0)
+		return -EFAULT;
+	input++;
+	if (get_user(fd, input) != 0)
 		return -EFAULT;
 
-	if (on) {
+	mutex_lock(&lguest_lock);
-		cpu->break_out = 1;
+	err = add_eventfd(lg, addr, fd);
-		/* Pop it out of the Guest (may be running on different CPU) */
+	mutex_unlock(&lguest_lock);
-		wake_up_process(cpu->tsk);
+
-		/* Wait for them to reset it */
-		return wait_event_interruptible(cpu->break_wq, !cpu->break_out);
-	} else {
-		cpu->break_out = 0;
-		wake_up(&cpu->break_wq);
 	return 0;
-	}
 }
 
 /*L:050 Sending an interrupt is done by writing LHREQ_IRQ and an interrupt
@@ -45,9 +96,8 @@ static int user_send_irq(struct lg_cpu *cpu, const unsigned long __user *input)
 		return -EFAULT;
 	if (irq >= LGUEST_IRQS)
 		return -EINVAL;
-	/* Next time the Guest runs, the core code will see if it can deliver
+
-	 * this interrupt. */
+	set_interrupt(cpu, irq);
-	set_bit(irq, cpu->irqs_pending);
 	return 0;
 }
 
@@ -126,9 +176,6 @@ static int lg_cpu_start(struct lg_cpu *cpu, unsigned id, unsigned long start_ip)
 	 * address. */
 	lguest_arch_setup_regs(cpu, start_ip);
 
-	/* Initialize the queue for the Waker to wait on */
-	init_waitqueue_head(&cpu->break_wq);
-
 	/* We keep a pointer to the Launcher task (ie. current task) for when
 	 * other Guests want to wake this one (eg. console input). */
 	cpu->tsk = current;
@@ -185,6 +232,13 @@ static int initialize(struct file *file, const unsigned long __user *input)
 		goto unlock;
 	}
 
+	lg->eventfds = kmalloc(sizeof(*lg->eventfds), GFP_KERNEL);
+	if (!lg->eventfds) {
+		err = -ENOMEM;
+		goto free_lg;
+	}
+	lg->eventfds->num = 0;
+
 	/* Populate the easy fields of our "struct lguest" */
 	lg->mem_base = (void __user *)args[0];
 	lg->pfn_limit = args[1];
@@ -192,7 +246,7 @@ static int initialize(struct file *file, const unsigned long __user *input)
 	/* This is the first cpu (cpu 0) and it will start booting at args[2] */
 	err = lg_cpu_start(&lg->cpus[0], 0, args[2]);
 	if (err)
-		goto release_guest;
+		goto free_eventfds;
 
 	/* Initialize the Guest's shadow page tables, using the toplevel
 	 * address the Launcher gave us.  This allocates memory, so can fail. */
@@ -211,7 +265,9 @@ static int initialize(struct file *file, const unsigned long __user *input)
 free_regs:
 	/* FIXME: This should be in free_vcpu */
 	free_page(lg->cpus[0].regs_page);
-release_guest:
+free_eventfds:
+	kfree(lg->eventfds);
+free_lg:
 	kfree(lg);
 unlock:
 	mutex_unlock(&lguest_lock);
@@ -252,11 +308,6 @@ static ssize_t write(struct file *file, const char __user *in,
 		/* Once the Guest is dead, you can only read() why it died. */
 		if (lg->dead)
 			return -ENOENT;
-
-		/* If you're not the task which owns the Guest, all you can do
-		 * is break the Launcher out of running the Guest. */
-		if (current != cpu->tsk && req != LHREQ_BREAK)
-			return -EPERM;
 	}
 
 	switch (req) {
@@ -264,8 +315,8 @@ static ssize_t write(struct file *file, const char __user *in,
 		return initialize(file, input);
 	case LHREQ_IRQ:
 		return user_send_irq(cpu, input);
-	case LHREQ_BREAK:
+	case LHREQ_EVENTFD:
-		return break_guest_out(cpu, input);
+		return attach_eventfd(lg, input);
 	default:
 		return -EINVAL;
 	}
@@ -303,6 +354,12 @@ static int close(struct inode *inode, struct file *file)
 		 * the Launcher's memory management structure. */
 		mmput(lg->cpus[i].mm);
 	}
+
+	/* Release any eventfds they registered. */
+	for (i = 0; i < lg->eventfds->num; i++)
+		fput(lg->eventfds->map[i].event);
+	kfree(lg->eventfds);
+
 	/* If lg->dead doesn't contain an error code it will be NULL or a
 	 * kmalloc()ed string, either of which is ok to hand to kfree(). */
 	if (!IS_ERR(lg->dead))

drivers/lguest/page_tables.c

@@ -53,6 +53,17 @@
  * page.  */
 #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1)
 
+/* For PAE we need the PMD index as well. We use the last 2MB, so we
+ * will need the last pmd entry of the last pmd page.  */
+#ifdef CONFIG_X86_PAE
+#define SWITCHER_PMD_INDEX 	(PTRS_PER_PMD - 1)
+#define RESERVE_MEM 		2U
+#define CHECK_GPGD_MASK		_PAGE_PRESENT
+#else
+#define RESERVE_MEM 		4U
+#define CHECK_GPGD_MASK		_PAGE_TABLE
+#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. */
@@ -73,24 +84,59 @@ 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
@@ -101,12 +147,32 @@ static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr)
 	return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t);
 }
 
-static unsigned long gpte_addr(pgd_t gpgd, unsigned long vaddr)
+#ifdef CONFIG_X86_PAE
+static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr)
 {
 	unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT;
 	BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT));
-	return gpage + ((vaddr>>PAGE_SHIFT) % PTRS_PER_PTE) * sizeof(pte_t);
+	return gpage + pmd_index(vaddr) * sizeof(pmd_t);
 }
+
+static unsigned long gpte_addr(struct lg_cpu *cpu,
+			       pmd_t gpmd, unsigned long vaddr)
+{
+	unsigned long gpage = pmd_pfn(gpmd) << PAGE_SHIFT;
+
+	BUG_ON(!(pmd_flags(gpmd) & _PAGE_PRESENT));
+	return gpage + pte_index(vaddr) * sizeof(pte_t);
+}
+#else
+static unsigned long gpte_addr(struct lg_cpu *cpu,
+				pgd_t gpgd, unsigned long vaddr)
+{
+	unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT;
+
+	BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT));
+	return gpage + pte_index(vaddr) * sizeof(pte_t);
+}
+#endif
 /*:*/
 
 /*M:014 get_pfn is slow: we could probably try to grab batches of pages here as
@@ -171,7 +237,7 @@ static void release_pte(pte_t pte)
 	/* Remember that get_user_pages_fast() took a reference to the page, in
 	 * get_pfn()?  We have to put it back now. */
 	if (pte_flags(pte) & _PAGE_PRESENT)
-		put_page(pfn_to_page(pte_pfn(pte)));
+		put_page(pte_page(pte));
 }
 /*:*/
 
@@ -184,11 +250,20 @@ static void check_gpte(struct lg_cpu *cpu, pte_t gpte)
 
 static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd)
 {
-	if ((pgd_flags(gpgd) & ~_PAGE_TABLE) ||
+	if ((pgd_flags(gpgd) & ~CHECK_GPGD_MASK) ||
 	   (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.
  *
@@ -207,6 +282,11 @@ 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);
 	/* Toplevel not present?  We can't map it in. */
@@ -228,12 +308,45 @@ bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode)
 		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));
+		set_pgd(spgd, __pgd(__pa(ptepage) | pgd_flags(gpgd)));
+	}
+
+#ifdef CONFIG_X86_PAE
+	gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t);
+	/* middle level not present?  We can't map it in. */
+	if (!(pmd_flags(gpmd) & _PAGE_PRESENT))
+		return false;
+
+	/* 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 false;
+		}
+
+		/* 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. */
+		native_set_pmd(spmd, __pmd(__pa(ptepage) | pmd_flags(gpmd)));
 	}
 
 	/* 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, gpmd, vaddr);
+#else
+	/* OK, now we look at the lower level in the Guest page table: keep its
+	 * address, because we might update it later. */
+	gpte_ptr = gpte_addr(cpu, gpgd, vaddr);
+#endif
 	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 +372,7 @@ 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);
@@ -273,7 +386,7 @@ bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode)
 		 * table entry, even if the Guest says it's writable.  That way
 		 * we will come back here when a write does actually occur, so
 		 * we can update the Guest's _PAGE_DIRTY flag. */
-		*spte = gpte_to_spte(cpu, pte_wrprotect(gpte), 0);
+		native_set_pte(spte, gpte_to_spte(cpu, pte_wrprotect(gpte), 0));
 
 	/* Finally, we write the Guest PTE entry back: we've set the
 	 * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */
@@ -301,14 +414,23 @@ 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 +444,43 @@ 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));
+	}
+}
+
+static void release_pgd(pgd_t *spgd)
+{
+	/* If the entry's not present, there's nothing to release. */
+	if (pgd_flags(*spgd) & _PAGE_PRESENT) {
+		unsigned int i;
+		pmd_t *pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
+
+		for (i = 0; i < PTRS_PER_PMD; i++)
+			release_pmd(&pmdpage[i]);
+
+		/* Now we can free the page of PMDs */
+		free_page((long)pmdpage);
+		/* And zero out the PGD entry so we never release it twice. */
+		set_pgd(spgd, __pgd(0));
+	}
+}
+
+#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) {
@@ -341,7 +498,7 @@ static void release_pgd(struct lguest *lg, pgd_t *spgd)
 		*spgd = __pgd(0);
 	}
 }
-
+#endif
 /*H:445 We saw flush_user_mappings() twice: once from the flush_user_mappings()
  * 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. */
@@ -350,7 +507,7 @@ 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,
@@ -369,7 +526,9 @@ 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);
 	/* Toplevel not present?  We can't map it in. */
@@ -378,7 +537,14 @@ unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr)
 		return -1UL;
 	}
 
-	gpte = lgread(cpu, gpte_addr(gpgd, vaddr), pte_t);
+#ifdef CONFIG_X86_PAE
+	gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t);
+	if (!(pmd_flags(gpmd) & _PAGE_PRESENT))
+		kill_guest(cpu, "Bad address %#lx", vaddr);
+	gpte = lgread(cpu, gpte_addr(cpu, gpmd, vaddr), pte_t);
+#else
+	gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t);
+#endif
 	if (!(pte_flags(gpte) & _PAGE_PRESENT))
 		kill_guest(cpu, "Bad address %#lx", vaddr);
 
@@ -405,6 +571,9 @@ static unsigned int new_pgdir(struct lg_cpu *cpu,
 			      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. */
@@ -416,11 +585,28 @@ static unsigned int new_pgdir(struct lg_cpu *cpu,
 		/* If the allocation fails, just keep using the one we have */
 		if (!cpu->lg->pgdirs[next].pgdir)
 			next = cpu->cpu_pgd;
-		else
+		else {
+#ifdef CONFIG_X86_PAE
+			/* 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);
+				set_pgd(cpu->lg->pgdirs[next].pgdir, __pgd(0));
+				next = cpu->cpu_pgd;
+			} else {
+				set_pgd(cpu->lg->pgdirs[next].pgdir +
+					SWITCHER_PGD_INDEX,
+					__pgd(__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
+			*blank_pgdir = 1;
+#endif
+		}
+	}
 	/* Record which Guest toplevel this shadows. */
 	cpu->lg->pgdirs[next].gpgdir = gpgdir;
 	/* Release all the non-kernel mappings. */
@@ -431,7 +617,7 @@ 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)
@@ -460,10 +646,25 @@ static void release_all_pagetables(struct lguest *lg)
 
 	/* 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) {
+#ifdef CONFIG_X86_PAE
+			pgd_t *spgd;
+			pmd_t *pmdpage;
+			unsigned int k;
+
+			/* 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
 			/* 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);
+		}
 }
 
 /* We also throw away everything when a Guest tells us it's changed a kernel
@@ -504,24 +705,37 @@ 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. */
+#ifdef CONFIG_X86_PAE
-		pte_t *spte = spte_addr(*spgd, vaddr);
+		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
+			/* If they're setting this entry as dirty or accessed,
-		 * as well put that entry they've given us in now.  This shaves
+			 * we might as well put that entry they've given us
-		 * 10% off a copy-on-write micro-benchmark. */
+			 * 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,
+				native_set_pte(spte,
-					     pte_flags(gpte) & _PAGE_DIRTY);
+						gpte_to_spte(cpu, gpte,
+						pte_flags(gpte) & _PAGE_DIRTY));
 			} else
-			/* Otherwise kill it and we can demand_page() it in
+				/* Otherwise kill it and we can demand_page()
-			 * later. */
+				 * it in later. */
-			*spte = __pte(0);
+				native_set_pte(spte, __pte(0));
+#ifdef CONFIG_X86_PAE
+		}
+#endif
 	}
 }
 
@@ -568,12 +782,10 @@ 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)
+void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx)
 {
 	int pgdir;
 
-	/* The kernel seems to try to initialize this early on: we ignore its
-	 * attempts to map over the Switcher. */
 	if (idx >= SWITCHER_PGD_INDEX)
 		return;
 
@@ -581,8 +793,14 @@ 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);
 }
+#ifdef CONFIG_X86_PAE
+void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx)
+{
+	guest_pagetable_clear_all(&lg->cpus[0]);
+}
+#endif
 
 /* Once we know how much memory we have we can construct simple identity
  * (which set virtual == physical) and linear mappings
@@ -596,8 +814,16 @@ 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
+	pmd_t __user *pmds;
+	unsigned int j;
+	pgd_t pgd;
+	pmd_t pmd;
+#else
+	unsigned int phys_linear;
+#endif
 
 	/* We have mapped_pages frames to map, so we need
 	 * linear_pages page tables to map them. */
@@ -610,6 +836,9 @@ 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. */
 	for (i = 0; i < mapped_pages; i++) {
@@ -621,6 +850,22 @@ static unsigned long setup_pagetables(struct lguest *lg,
 
 	/* The top level points to the linear page table pages above.
 	 * We setup the identity and linear mappings here. */
+#ifdef CONFIG_X86_PAE
+	for (i = j = 0; i < mapped_pages && j < PTRS_PER_PMD;
+	     i += PTRS_PER_PTE, j++) {
+		native_set_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;
+	}
+
+	set_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 +878,7 @@ 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. */
@@ -648,7 +894,10 @@ 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). */
 	if (copy_from_user(&mem, &boot->e820_map[0].size, sizeof(mem))
@@ -663,6 +912,15 @@ int init_guest_pagetable(struct lguest *lg)
 	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;
+
+	set_pgd(pgd + SWITCHER_PGD_INDEX,
+		__pgd(__pa(pmd_table) | _PAGE_PRESENT));
+#endif
 	lg->cpus[0].cpu_pgd = 0;
 	return 0;
 }
@@ -673,16 +931,23 @@ void page_table_guest_data_init(struct lg_cpu *cpu)
 	/* 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
+		/* We tell the Guest that it can't use the top 2 or 4 MB
-	     * addresses used by the Switcher. */
+		 * of virtual addresses used by the Switcher. */
-	    || put_user(4U*1024*1024, &cpu->lg->lguest_data->reserve_mem)
+		|| put_user(RESERVE_MEM * 1024 * 1024,
-	    || put_user(cpu->lg->pgdirs[0].gpgdir, &cpu->lg->lguest_data->pgdir))
+			&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. */
+#ifdef CONFIG_X86_PAE
+	if (pgd_index(cpu->lg->kernel_address) == SWITCHER_PGD_INDEX &&
+		pmd_index(cpu->lg->kernel_address) == SWITCHER_PMD_INDEX)
+#else
 	if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX)
+#endif
 		kill_guest(cpu, "bad kernel address %#lx",
 				 cpu->lg->kernel_address);
 }
@@ -708,16 +973,30 @@ 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;
+
+	native_set_pmd(&switcher_pmd, pfn_pmd(__pa(switcher_pte_page) >>
+		       PAGE_SHIFT, PAGE_KERNEL_EXEC));
+
+	pmd_table = __va(pgd_pfn(cpu->lg->
+			pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX])
+								<< PAGE_SHIFT);
+	native_set_pmd(&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);
+	switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL_EXEC);
 
 	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
 	 * page for this CPU is.  This is an optimization: when the Switcher
@@ -726,8 +1005,9 @@ void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages)
 	 * page is already mapped there, we don't have to copy them out
 	 * again. */
 	pfn = __pa(cpu->regs_page) >> PAGE_SHIFT;
-	regs_pte = pfn_pte(pfn, __pgprot(__PAGE_KERNEL));
+	native_set_pte(&regs_pte, pfn_pte(pfn, PAGE_KERNEL));
-	switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTRS_PER_PTE] = regs_pte;
+	native_set_pte(&switcher_pte_page[pte_index((unsigned long)pages)],
+			regs_pte);
 }
 /*:*/
 
@@ -752,21 +1032,21 @@ 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],
+		native_set_pte(&pte[i], mk_pte(switcher_page[i],
-				__pgprot(_PAGE_PRESENT|_PAGE_ACCESSED));
+				__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]),
+	native_set_pte(&pte[i], pfn_pte(page_to_pfn(switcher_page[i]),
-			 __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW));
+			 __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]),
+	native_set_pte(&pte[i+1], pfn_pte(page_to_pfn(switcher_page[i+1]),
-			   __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED));
+			   __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)));
 }
 
 /* We've made it through the page table code.  Perhaps our tired brains are

drivers/lguest/segments.c

@@ -150,7 +150,7 @@ void load_guest_gdt_entry(struct lg_cpu *cpu, u32 num, u32 lo, u32 hi)
 {
 	/* We assume the Guest has the same number of GDT entries as the
 	 * Host, otherwise we'd have to dynamically allocate the Guest GDT. */
-	if (num > ARRAY_SIZE(cpu->arch.gdt))
+	if (num >= ARRAY_SIZE(cpu->arch.gdt))
 		kill_guest(cpu, "too many gdt entries %i", num);
 
 	/* Set it up, then fix it. */

fs/eventfd.c

@@ -16,6 +16,7 @@
 #include <linux/anon_inodes.h>
 #include <linux/eventfd.h>
 #include <linux/syscalls.h>
+#include <linux/module.h>
 
 struct eventfd_ctx {
 	wait_queue_head_t wqh;
@@ -56,6 +57,7 @@ int eventfd_signal(struct file *file, int n)
 
 	return n;
 }
+EXPORT_SYMBOL_GPL(eventfd_signal);
 
 static int eventfd_release(struct inode *inode, struct file *file)
 {
@@ -197,6 +199,7 @@ struct file *eventfd_fget(int fd)
 
 	return file;
 }
+EXPORT_SYMBOL_GPL(eventfd_fget);
 
 SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags)
 {

include/linux/lguest.h

@@ -30,6 +30,10 @@ struct lguest_data
 	/* Wallclock time set by the Host. */
 	struct timespec time;
 
+	/* Interrupt pending set by the Host.  The Guest should do a hypercall
+	 * if it re-enables interrupts and sees this set (to X86_EFLAGS_IF). */
+	int irq_pending;
+
 	/* Async hypercall ring.  Instead of directly making hypercalls, we can
 	 * place them in here for processing the next time the Host wants.
 	 * This batching can be quite efficient. */

include/linux/lguest_launcher.h

@@ -57,7 +57,8 @@ enum lguest_req
 	LHREQ_INITIALIZE, /* + base, pfnlimit, start */
 	LHREQ_GETDMA, /* No longer used */
 	LHREQ_IRQ, /* + irq */
-	LHREQ_BREAK, /* + on/off flag (on blocks until someone does off) */
+	LHREQ_BREAK, /* No longer used */
+	LHREQ_EVENTFD, /* + address, fd. */
 };
 
 /* The alignment to use between consumer and producer parts of vring.

kernel/sched.c

@@ -2192,6 +2192,7 @@ void kick_process(struct task_struct *p)
 		smp_send_reschedule(cpu);
 	preempt_enable();
 }
+EXPORT_SYMBOL_GPL(kick_process);
 
 /*
  * Return a low guess at the load of a migration-source cpu weighted