[lkml]   [2007]   [Aug]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
Subject[PATCH] Fix preemptible lazy mode bug
I recently sent off a fix for lazy vmalloc faults which can happen under 
paravirt when lazy mode is enabled. Unfortunately, I jumped the gun a
bit on fixing this. I neglected to notice that since the new call to
flush the MMU update queue is called from the page fault handler, it can
be pre-empted. Both VMI and Xen use per-cpu variables to track lazy
mode state, as all previous calls to set, disable, or flush lazy mode
happened from a non-preemptable state.

I have no idea how to convincingly produce the problem, as generating a
kernel pre-emption at the required point is, um, difficult, but it is
most certainly a real possibility, and potentially more likely than the
bug I fixed originally.

Rusty, you may have to modify lguest code if you use lazy mode and rely
on per-cpu variables during the callout for paravirt_ops.set_lazy_mode.

I have tested as best as I can, and am trying to write a suite destined
for LTP which will help catch and debug these issues.

Since set_lazy_mode(LAZY_MODE_FLUSH) is now called from the page fault
handler, it can potentially happen in a preemptible state. We therefore
must make all lazy mode paravirt-ops handlers non-preemptible.

Signed-off-by: Zachary Amsden <zamsden@mysore.(none)>
arch/i386/kernel/vmi.c | 14 ++++++++++----
arch/i386/xen/enlighten.c | 4 +++-
2 files changed, 13 insertions(+), 5 deletions(-)

diff --git a/arch/i386/kernel/vmi.c b/arch/i386/kernel/vmi.c
index 18673e0..9e669cb 100644
--- a/arch/i386/kernel/vmi.c
+++ b/arch/i386/kernel/vmi.c
@@ -555,21 +555,27 @@ vmi_startup_ipi_hook(int phys_apicid, unsigned long start_eip,
static void vmi_set_lazy_mode(enum paravirt_lazy_mode mode)
static DEFINE_PER_CPU(enum paravirt_lazy_mode, lazy_mode);
+ int cpu;
+ enum paravirt_lazy_mode cur_mode;

if (!vmi_ops.set_lazy_mode)

+ cpu = get_cpu();
+ cur_mode = per_cpu(lazy_mode, cpu);
/* Modes should never nest or overlap */
- BUG_ON(__get_cpu_var(lazy_mode) && !(mode == PARAVIRT_LAZY_NONE ||
+ BUG_ON(cur_mode && !(mode == PARAVIRT_LAZY_NONE ||

if (mode == PARAVIRT_LAZY_FLUSH) {
- vmi_ops.set_lazy_mode(__get_cpu_var(lazy_mode));
+ vmi_ops.set_lazy_mode(cur_mode);
} else {
- __get_cpu_var(lazy_mode) = mode;
+ per_cpu(lazy_mode, cpu) = mode;
+ put_cpu();

static inline int __init check_vmi_rom(struct vrom_header *rom)
diff --git a/arch/i386/xen/enlighten.c b/arch/i386/xen/enlighten.c
index f0c3751..2dafb8a 100644
--- a/arch/i386/xen/enlighten.c
+++ b/arch/i386/xen/enlighten.c
@@ -251,7 +251,7 @@ static void xen_halt(void)

static void xen_set_lazy_mode(enum paravirt_lazy_mode mode)
- BUG_ON(preemptible());
+ get_cpu();

switch (mode) {
@@ -267,11 +267,13 @@ static void xen_set_lazy_mode(enum paravirt_lazy_mode mode)
/* flush if necessary, but don't change state */
if (x86_read_percpu(xen_lazy_mode) != PARAVIRT_LAZY_NONE)
+ put_cpu();

x86_write_percpu(xen_lazy_mode, mode);
+ put_cpu();

static unsigned long xen_store_tr(void)
 \ /
  Last update: 2007-08-24 07:59    [W:0.084 / U:7.608 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site