On Thu, May 16, 2019 at 6:18 PM Filipe Manana wrote: > > On Thu, May 16, 2019 at 5:59 PM Theodore Ts'o wrote: > > > > On Thu, May 16, 2019 at 10:54:57AM +0100, Filipe Manana wrote: > > > > > > Haven't tried ext4 with 1 process only (instead of 4), but I can try > > > to see if it happens without concurrency as well. > > > > How many CPU's and how much memory were you using? And I assume this > > was using KVM/QEMU? How was it configured? > > Yep, kvm and qemu (3.0.0). The qemu config: > > https://pastebin.com/KNigeXXq > > TEST_DEV is the drive with ID "drive1" and SCRATCH_DEV is the drive > with ID "drive2". > > The host has: > > Intel(R) Core(TM) i7-8700 CPU @ 3.20GHz > 64Gb of ram > crappy seagate hdd: > > Device Model: ST3000DM008-2DM166 > Serial Number: Z5053T2R > LU WWN Device Id: 5 000c50 0a46f7ecb > Firmware Version: CC26 > User Capacity: 3,000,592,982,016 bytes [3,00 TB] > Sector Sizes: 512 bytes logical, 4096 bytes physical > Rotation Rate: 7200 rpm > Form Factor: 3.5 inches > Device is: Not in smartctl database [for details use: -P showall] > ATA Version is: ACS-2, ACS-3 T13/2161-D revision 3b > SATA Version is: SATA 3.1, 6.0 Gb/s (current: 6.0 Gb/s) > > It hosts 3 qemu instances, all with the same configuration. > > I left the test running earlier today for about 1 hour on ext4 with > only 1 fsstress process. Didn't manage to reproduce. > With 4 or more processes, those journal checksum failures happen sporadically. > I can leave it running with 1 process during this evening and see what > we get here, if it happens with 1 process, it should be trivial to > reproduce anywhere. Ok, so I left it running overnight, for 17 000+ iterations. It failed 102 times with that journal corruption. I changed the test to randomize the number of fsstress processes between 1 to 8. I'm attaching here the logs (.full, .out.bad and dmesg files) in case you are interested in the seed values for fsstress. So the test does now: (...) procs=$(( (RANDOM % 8) + 1 )) args=`_scale_fsstress_args -p $procs -n 100 $FSSTRESS_AVOID -d $SCRATCH_MNT/test` args="$args -f mknod=0 -f symlink=0" echo "Running fsstress with arguments: $args" >>$seqres.full (...) I verified no failures happened with only 1 process, and the more processes are used, the more likely it is to hit the issue: $ egrep -r 'Running fsstress with arguments: -p' . | cut -d ' ' -f 6 | perl -n -e 'use Statistics::Histogram; @data = <>; chomp @data; print get_histogram(\@data);' Count: 102 Range: 2.000 - 8.000; Mean: 5.598; Median: 6.000; Stddev: 1.831 Percentiles: 90th: 8.000; 95th: 8.000; 99th: 8.000 2.000 - 2.348: 5 ############## 2.348 - 3.171: 13 #################################### 3.171 - 4.196: 12 ################################# 4.196 - 5.473: 15 ########################################## 5.473 - 6.225: 19 ##################################################### 6.225 - 7.064: 19 ##################################################### 7.064 - 8.000: 19 ##################################################### And verified picking one of the failing seeds, such as 1557322233 for 2 processes, and running the test with that seed for 10 times didn't reproduce, so it indeed seems to be some race causing the journal corruption. Forgot previously, but my kernel config in case it helps: https://pastebin.com/LKvRcAW1 Thanks. > > > > > Thanks, > > > > - Ted