numam-spdk/test/scheduler/interrupt.sh
Michal Berger 049d587e50 test/scheduler: Add basic tests for verifying interrupt mode
Also, move all the functions wrapping the scheduler app into
common.sh.

Signed-off-by: Michal Berger <michalx.berger@intel.com>
Change-Id: I63c1002f867128851cff04a035c6c51c50e10383
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/6053
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
2021-01-25 20:37:50 +00:00

134 lines
4.2 KiB
Bash
Executable File

#!/usr/bin/env bash
testdir=$(readlink -f "$(dirname "$0")")
rootdir=$(readlink -f "$testdir/../../")
source "$rootdir/test/common/autotest_common.sh"
source "$testdir/common.sh"
trap 'killprocess "$spdk_pid"' EXIT
declare -a cpus=()
declare -a cpus_to_collect=()
fold_list_onto_array cpus $(parse_cpu_list <(echo "$spdk_cpus_csv"))
# Normalize the indexes
cpus=("${cpus[@]}")
collect_cpu_stat() {
local list=$1
local stat=$2
for cpu in "${cpus_to_collect[@]}"; do
eval "${list}[cpu]=\$(get_cpu_stat $cpu $stat)"
done
}
collect_cpu_idle() {
xtrace_disable
local sample_time=${1:-5} samples=0
local cpu bool inc user_hz
# idle scales to USER_HZ so we use that in order to determine the expected
# value it should have been increased to (more or less).
user_hz=100
# Expected increase of the idle stat
inc=$((user_hz * sample_time))
bool[0]="not" bool[1]="is"
init_idle_samples=() idle_samples=() is_idle=()
collect_cpu_stat init_idle_samples idle
printf 'Collecting cpu idle stats (cpus: %s) for %u seconds...\n' \
"${cpus_to_collect[*]}" "$sample_time"
while ((++samples <= sample_time)) && sleep 1s; do
collect_cpu_stat idle_samples idle
done
for cpu in "${!idle_samples[@]}"; do
# We start to collect after the spdk app is initialized hence if the interrupt
# mode is not working as expected, the idle time of given cpu will not have a
# chance to increase. If it does work correctly, then it should change even for
# a fraction, depending on how much time we spent on collecting this data.
# If idle time is over 70% of expected increase then we consider this cpu as
# idle. This is done in order to take into consideration time window the app
# needs to actually spin up|down the cpu. It's also taken for granted that
# there is no extra load on the target cpus which may be coming from other
# processes.
if ((idle_samples[cpu] > init_idle_samples[cpu] + (inc * 70 / 100))); then
is_idle[cpu]=1
else
is_idle[cpu]=0
fi
printf 'cpu%u %s idle (%u %u)\n' \
"$cpu" "${bool[is_idle[cpu]]}" "${init_idle_samples[cpu]}" "${idle_samples[cpu]}"
done
xtrace_restore
}
interrupt() {
local busy_cpus
local cpu thread
local reactor_framework
cpus_to_collect=("${cpus[@]}")
collect_cpu_idle
# Standard scenario - spdk app is idle, all cpus, except the main cpu, should be
# switched into interrupt mode. main cpu should remain busy, all remaining cpus
# should become idle.
for cpu in "${!is_idle[@]}"; do
if ((cpu == spdk_main_core)); then
((is_idle[cpu] == 0)) # main cpu must not be idle
fi
if ((cpu != spdk_main_core)); then
((is_idle[cpu] == 1)) # all cpus except the main cpu must be idle
fi
done
# While cpus are still idle, verify that each cpu, except the main cpu, has no threads assigned
reactor_framework=$(rpc_cmd framework_get_reactors | jq -r '.reactors[]')
for cpu in "${cpus[@]:1}"; do
[[ -z $(jq -r "select(.lcore == $cpu) | .lw_threads[].id" <<< "$reactor_framework") ]]
done
# select 3 cpus except the main one
busy_cpus=("${cpus[@]:1:3}") threads=()
# Create busy thread on each of the selected cpus, then verify if given cpu has become busy
# and that newly created thread was assigned to a proper lcore.
for cpu in "${busy_cpus[@]}"; do
threads[cpu]=$(create_thread -n "thread$cpu" -m "$(mask_cpus "$cpu")" -a 100) cpus_to_collect=("$cpu")
collect_cpu_idle
((is_idle[cpu] == 0))
reactor_framework=$(rpc_cmd framework_get_reactors | jq -r '.reactors[]')
[[ -n $(jq -r "select(.lcore == $cpu) | .lw_threads[] | select(.name == \"thread$cpu\")" <<< "$reactor_framework") ]]
done
# Make all the threads idle and verify if their cpus have become idle as well and if they were
# moved away out of their lcores.
for cpu in "${!threads[@]}"; do
active_thread "${threads[cpu]}" 0
cpus_to_collect=("$cpu")
# Give some extra time for the cpu to spin down
collect_cpu_idle 10
((is_idle[cpu] == 1))
reactor_framework=$(rpc_cmd framework_get_reactors | jq -r '.reactors[]')
[[ -z $(jq -r "select(.lcore == $cpu) | .lw_threads[].id" <<< "$reactor_framework") ]]
done
for cpu in "${!threads[@]}"; do
destroy_thread "${threads[cpu]}"
done
}
exec_under_dynamic_scheduler "$scheduler" -m "$spdk_cpusmask" --main-core "$spdk_main_core"
interrupt