Skip to content

Commit

Permalink
sched: do not account for NMIs
Browse files Browse the repository at this point in the history
Impact: avoid corruption in system time accounting

Martin Schwidefsky told me that there was an issue with NMIs and
system accounting. The problem is that the accounting code is
not reentrant, and if an NMI goes off after an interrupt it can
corrupt the accounting.

For now, the best we can do is to treat NMIs like SMIs and they
are not accounted for.

This patch changes nmi_enter to not call __irq_enter and to do
the preempt-count and tracing calls directly.

Signed-off-by: Steven Rostedt <srostedt@redhat.com>
  • Loading branch information
Steven Rostedt committed Feb 12, 2009
1 parent 5a5fb7d commit 2a7b8df
Showing 1 changed file with 16 additions and 16 deletions.
32 changes: 16 additions & 16 deletions include/linux/hardirq.h
Original file line number Diff line number Diff line change
Expand Up @@ -175,24 +175,24 @@ extern void irq_enter(void);
*/
extern void irq_exit(void);

#define nmi_enter() \
do { \
ftrace_nmi_enter(); \
BUG_ON(in_nmi()); \
add_preempt_count(NMI_OFFSET); \
lockdep_off(); \
rcu_nmi_enter(); \
__irq_enter(); \
#define nmi_enter() \
do { \
ftrace_nmi_enter(); \
BUG_ON(in_nmi()); \
add_preempt_count(NMI_OFFSET + HARDIRQ_OFFSET); \
lockdep_off(); \
rcu_nmi_enter(); \
trace_hardirq_enter(); \
} while (0)

#define nmi_exit() \
do { \
__irq_exit(); \
rcu_nmi_exit(); \
lockdep_on(); \
BUG_ON(!in_nmi()); \
sub_preempt_count(NMI_OFFSET); \
ftrace_nmi_exit(); \
#define nmi_exit() \
do { \
trace_hardirq_exit(); \
rcu_nmi_exit(); \
lockdep_on(); \
BUG_ON(!in_nmi()); \
sub_preempt_count(NMI_OFFSET + HARDIRQ_OFFSET); \
ftrace_nmi_exit(); \
} while (0)

#endif /* LINUX_HARDIRQ_H */

0 comments on commit 2a7b8df

Please sign in to comment.