From f58d7dbdab7076d1ff83375fdbc010445261c0cd Mon Sep 17 00:00:00 2001 From: Chris Evich Date: Mon, 13 Jun 2022 12:12:13 -0400 Subject: Cirrus: Fix elevator workaround multi-cloud support In order to support execution on various non-GCP cloud environments, the BFQ scheduler workaround needs updating. Previously it assumed the root disk was always `/dev/sda`. With the addition of new clouds (AWS) and different environment types, the assumption is not always valid. Update the workaround to take care in looking up the block device where '/' comes from. Also update the scheduler to 'none', as all modern clouds already have highly optimized underlying storage configurations. There's no reason to complicate I/O paths further by hard-coding specific scheduler(s) for all environment types. Signed-off-by: Chris Evich --- contrib/cirrus/setup_environment.sh | 27 ++++++++++++++++++++++----- 1 file changed, 22 insertions(+), 5 deletions(-) (limited to 'contrib') diff --git a/contrib/cirrus/setup_environment.sh b/contrib/cirrus/setup_environment.sh index 9bd35bd06..fa0a40991 100755 --- a/contrib/cirrus/setup_environment.sh +++ b/contrib/cirrus/setup_environment.sh @@ -99,11 +99,28 @@ esac if ((CONTAINER==0)); then # Not yet running inside a container # Discovered reemergence of BFQ scheduler bug in kernel 5.8.12-200 # which causes a kernel panic when system is under heavy I/O load. - # Previously discovered in F32beta and confirmed fixed. It's been - # observed in F31 kernels as well. Deploy workaround for all VMs - # to ensure a more stable I/O scheduler (elevator). - echo "mq-deadline" > /sys/block/sda/queue/scheduler - warn "I/O scheduler: $(cat /sys/block/sda/queue/scheduler)" + # Disable the I/O scheduler (a.k.a. elevator) for all environments, + # leaving optimization up to underlying storage infrastructure. + testfs="/" # mountpoint that experiences the most I/O during testing + msg "Querying block device owning partition hosting the '$testfs' filesystem" + # Need --nofsroot b/c btrfs appends subvolume label to `source` name + testdev=$(findmnt --canonicalize --noheadings --nofsroot \ + --output source --mountpoint $testfs) + msg " found partition: '$testdev'" + testdisk=$(lsblk --noheadings --output pkname --paths $testdev) + msg " found block dev: '$testdisk'" + testsched="/sys/block/$(basename $testdisk)/queue/scheduler" + if [[ -n "$testdev" ]] && [[ -n "$testdisk" ]] && [[ -e "$testsched" ]]; then + msg " Found active I/O scheduler: $(cat $testsched)" + if [[ ! "$(<$testsched)" =~ \[none\] ]]; then + msg " Disabling elevator for '$testsched'" + echo "none" > "$testsched" + else + msg " Elevator already disabled" + fi + else + warn "Sys node for elevator doesn't exist: '$testsched'" + fi fi # Which distribution are we testing on. -- cgit v1.2.3-54-g00ecf