We *used* to not any error handling as part of the block layer's
*add_disk() paths since the code's inception. Fortunately that's now
history, but the only piece of code changes that were dropped from
that effort was error injection [0], this was since Hannes noted that
we may want to discuss if the approach is the best we can do.
I looked into this and indeed the BFP method to do error injection
is a viable alternative [1]. However this was even further generalized
from kprobes and all one now needs is to sprinkle ALLOW_ERROR_INJECTION()
on calls we went to enable error injection for. This makes things much
easier, instead of having to have a kernel bpf program to load
load_bpf_file() and then to run that and specify the comands you want
on the shell you can now just use something as simple as just shell:
-------------------------------------------------------------------
#!/bin/bash
rm -f testfile.img
dd if=/dev/zero of=testfile.img bs=1M seek=1000 count=1
DEVICE=$(losetup --show -f testfile.img)
mkfs.btrfs -f $DEVICE
mkdir -p tmpmnt
FAILTYPE=fail_function
FAILFUNC=open_ctree
echo $FAILFUNC > /sys/kernel/debug/$FAILTYPE/inject
echo -12 > /sys/kernel/debug/$FAILTYPE/$FAILFUNC/retval
echo N > /sys/kernel/debug/$FAILTYPE/task-filter
echo 100 > /sys/kernel/debug/$FAILTYPE/probability
echo 0 > /sys/kernel/debug/$FAILTYPE/interval
echo -1 > /sys/kernel/debug/$FAILTYPE/times
echo 0 > /sys/kernel/debug/$FAILTYPE/space
echo 1 > /sys/kernel/debug/$FAILTYPE/verbose
mount -t btrfs $DEVICE tmpmnt
if [ $? -ne 0 ]
then
echo "SUCCESS!"
else
echo "FAILED!"
umount tmpmnt
fi
echo > /sys/kernel/debug/$FAILTYPE/inject
rmdir tmpmnt
losetup -d $DEVICE
rm testfile.img
-------------------------------------------------------------------
This seems to be much more adaptable to what we do in blktests and
fstests. So before I go forward with adding error injection for the
block layer (only one user) or fs (only btrfs uses this so far), I think
it would be prudent for us to socialize if this *is* the scalable
strategy we'd like to see moving forward. If not then LSFMM seems like a
good place to iron out any possible kinks or concerns folks might have.
If we already have consensus and this is *the* way to go then I'll just
go forward and start adding some knobs / tests for this.
Perhaps the only issue I can think of is that you need a kernel
which enables error injection, and so production kernels would not
cut it.
Thoughts?
[0] https://lkml.kernel.org/r/20210512064629.13899-9-mcgrof@xxxxxxxxxx
[1] https://lwn.net/Articles/740146/
[2] https://patchwork.ozlabs.org/project/netdev/patch/151563182380.628.2420967932180154822.stgit@devbox/
Luis
