out of order execution of shell commands
we have a weird bug where the open of the shell output redirection happens before the mkdir command returns.
As a result, the redirection will fail. Unexpected data loss.
This bug was seen on ppc64 (dual ppc970) and also on i386, single cpu 3GHz xeon with smp kernel. Running kernel was our SLES9 2.6.5. It did also happen on ppc64 with 2.6.9-rc2-mm4, but I havent verified that. It was not seen with just one script running, but with 2 (or more).
The inner loop fails after 2-3 hours, sometimes earlier, sometimes never. Some tracing shows:
sys32_open(2076) fs-stress-olh.s(2736):c1,j3193150 S /dev/zero sys32_open(2076) fs-stress-olh.s(2736):c1,j3193154 S 01/11/59/8/data sys32_open(2076) fs-stress-olh.s(2736):c1,j3193157 S ./01/11/59/9/DIRNAME sys_mkdir(1696) fs-stress-olh.s(2736):c1,j3193162 E 01/11/59/9
That means, a gap of up 5 jiffies betwen the start of the open syscall and the end of the mkdir syscall.
In this case, everything happend on cpu1.
The dd command does never fail.
If I add more printk (start/end of sys_open, sys_mkdir), the bug will disappear.
A quick look at the bash2 sources shows a wait(-1,..), I believe the bash does nothing unusual.
Any ideas whats going on here, and how to debug it further?
testcase:
fs-stress-olh.sh
#!/bin/bash
export PATH=/bin/echo $$ > /dev/shm/bug42232-pid-$$/bin/mkdir stress-root
cd stress-root
/bin/echo Setting up ...
for i in 0 1 2 3 4 5 6 7 8 9; do
for j in 0 1 2 3 4 5 6 7 8 9; do
for k in 0 1 2 3 4 5 6 7 8 9; do
/bin/echo -n "$i$j$k " >> onethousand
done
/bin/echo -n "$i$j " >> onehundret done
/bin/echo -n "$i " >> ten
done/bin/echo -n "0 1 2 3 4 5 6 7 8 9" > onethousand
for h in 0 1 ; do
for i in /bin/cat onethousand; do
/bin/mkdir $h$i
/bin/echo $h$i > ./$h$i/DIRNAME
for j in `/bin/cat onehundret`; do
/bin/mkdir $h$i/$j
/bin/echo $h$i/$j > ./$h$i/$j/DIRNAME
for k in `/bin/cat onehundret`; do
/bin/mkdir $h$i/$j/$k
/bin/echo $h$i/$j/$k > ./$h$i/$j/$k/DIRNAME
for l in `/bin/cat ten`; do
/bin/mkdir $h$i/$j/$k/$l
/bin/echo $h$i/$j/$k/$l > ./$h$i/$j/$k/$l/DIRNAME || /bin/kill -STOP `/bin/cat /dev/shm/bug42232-pid-*`
/bin/dd if=/dev/zero of=$h$i/$j/$k/$l/data bs=1024 count=1 &> /dev/zero
done
done
done done
done
cd ../bin/rm -rf stress-root /dev/shm/bug42232-pid-$$
start_test.sh
#!/bin/bash
set -ex
cd /bug42232
test -x ~olh/bugs/42232/fs-stress-olh.sh
numcpus=grep -Ec '^cpu[0-9]' /proc/stat
for i in seq 1 $numcpus
do
(rm -rf $i ; mkdir -v $i; cd $i || exit 1 ; pwd ; nohup sh -c 'while date ; do ~olh/bugs/42232/fs-stress-olh.sh ; done' < /dev/null > logfile 2>&1 & ) done
sleep $numcpus
( tail --follow=name --retry --lines=42 ?/logfile & )
Signature · 3 lines
USB is for mice, FireWire is for men! sUse lINUX ag, nÜRNBERG
Quoted 4 lines · Olaf Hering
we have a weird bug where the open of the shell output redirection happens before the mkdir command returns.
As a result, the redirection will fail. Unexpected data loss.
...
Quoted 5 lines
for l in `/bin/cat ten`; do
/bin/mkdir $h$i/$j/$k/$l
/bin/echo $h$i/$j/$k/$l > ./$h$i/$j/$k/$l/DIRNAME || /bin/kill -STOP `/bin/cat /dev/shm/bug42232-pid-*`
/bin/dd if=/dev/zero of=$h$i/$j/$k/$l/data bs=1024 count=1 &> /dev/zero
done...
looks like the dd and the mkdir will get the same pid, and bash can not deal with it.
Signature · 3 lines
USB is for mice, FireWire is for men! sUse lINUX ag, nÜRNBERG