From: Joshua Wise <[email protected]>
Background:
When a userspace application wants to know about machine check events, it
opens /dev/mcelog and does a read(). Usually, we found that this interface
works well, but in some cases, when the system was taking large numbers of
machine check exceptions, the read() would hang. The system would output a
soft-lockup warning, and the daemon reading from /dev/mcelog would suck up
as much of a single CPU as it could spinning in system space.
Description:
This patch fixes this bug. In particular, there was a "continue" inside a
timeout loop that presumably was intended to break out of the outer loop,
but instead caused the inner loop to continue. This patch also makes the
condition for the break-out a little more evident by changing a
!time_before to a time_after_eq.
Result:
The read() no longer hangs in this test case.
Testing:
On my system, I could replicate the bug with the following command:
# for i in `seq 15000`; do ./inject_sbe.sh; done
where inject_sbe.sh contains commands to inject a single-bit error into the
next memory write transaction.
Patch:
This patch is against git f1518a088bde6aea49e7c472ed6ab96178fcba3e.
Signed-off-by: Joshua Wise <[email protected]>
Signed-off-by: Tim Hockin <[email protected]>
--
diff --git a/arch/x86_64/kernel/mce.c b/arch/x86_64/kernel/mce.c
index a14375d..aa83023 100644
--- a/arch/x86_64/kernel/mce.c
+++ b/arch/x86_64/kernel/mce.c
@@ -497,15 +497,17 @@ static ssize_t mce_read(struct file *fil
for (i = 0; i < next; i++) {
unsigned long start = jiffies;
while (!mcelog.entry[i].finished) {
- if (!time_before(jiffies, start + 2)) {
+ if (time_after_eq(jiffies, start + 2)) {
memset(mcelog.entry + i,0, sizeof(struct mce));
- continue;
+ goto timeout;
}
cpu_relax();
}
smp_rmb();
err |= copy_to_user(buf, mcelog.entry + i, sizeof(struct mce));
buf += sizeof(struct mce);
+ timeout:
+ ;
}
memset(mcelog.entry, 0, next * sizeof(struct mce));
On Thu, 21 Jun 2007 17:13:49 -0700 (PDT) Joshua Wise wrote:
> Testing:
> On my system, I could replicate the bug with the following command:
> # for i in `seq 15000`; do ./inject_sbe.sh; done
> where inject_sbe.sh contains commands to inject a single-bit error into the
> next memory write transaction.
Hi,
Does inject_sbe.sh require special hardware? If so, what hardware
is it?
Thanks,
---
~Randy
*** Remember to use Documentation/SubmitChecklist when testing your code ***