From 58a89a1cc0760daa515f58da9bb8b167f01044bb Mon Sep 17 00:00:00 2001
From: Daniel Sampliner <samplinerD@gmail.com>
Date: Sat, 1 Feb 2025 17:51:46 -0500
Subject: [PATCH] fix: Parallelism CPU calculation inside Kubernetes and Docker
 with limits (#799)

The value of /sys/fs/cgroup/cpu/cpu.cfs_quota_us is not in milliseconds
and cannot be simply divided by 1000 to determine the CPU limit. As per
kernel documentation[^1], the cpu limit can be determined by dividing
that value by /sys/fs/cgroup/cpu/cpu.cfs_period_us.

[^1]: https://docs.kernel.org/scheduler/sched-bwc.html

---------

Co-authored-by: George L. Yermulnik <yz@yz.kiev.ua>
Co-authored-by: MaxymVlasov <MaxymVlasov@users.noreply.github.com>
---
 hooks/_common.sh | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)

diff --git a/hooks/_common.sh b/hooks/_common.sh
index fc0c930a7..e9defff52 100644
--- a/hooks/_common.sh
+++ b/hooks/_common.sh
@@ -188,6 +188,11 @@ function common::is_hook_run_on_whole_repo {
 
 #######################################################################
 # Get the number of CPU logical cores available for pre-commit to use
+#
+# CPU quota should be calculated as `cpu.cfs_quota_us / cpu.cfs_period_us`
+# For K8s see: https://docs.kernel.org/scheduler/sched-bwc.html
+# For Docker see: https://docs.docker.com/engine/containers/resource_constraints/#configure-the-default-cfs-scheduler
+#
 # Arguments:
 #  parallelism_ci_cpu_cores (string) Used in edge cases when number of
 #    CPU cores can't be derived automatically
@@ -197,14 +202,15 @@ function common::is_hook_run_on_whole_repo {
 function common::get_cpu_num {
   local -r parallelism_ci_cpu_cores=$1
 
-  local millicpu
+  local cpu_quota cpu_period cpu_num
 
   if [[ -f /sys/fs/cgroup/cpu/cpu.cfs_quota_us &&
     ! -f /proc/sys/fs/binfmt_misc/WSLInterop ]]; then # WSL have cfs_quota_us, but WSL should be checked as usual Linux host
     # Inside K8s pod or DinD in K8s
-    millicpu=$(< /sys/fs/cgroup/cpu/cpu.cfs_quota_us)
+    cpu_quota=$(< /sys/fs/cgroup/cpu/cpu.cfs_quota_us)
+    cpu_period=$(cat /sys/fs/cgroup/cpu/cpu.cfs_period_us 2> /dev/null || echo "$cpu_quota")
 
-    if [[ $millicpu -eq -1 ]]; then
+    if [[ $cpu_quota -eq -1 || $cpu_period -lt 1 ]]; then
       # K8s no limits or in DinD
       if [[ -n $parallelism_ci_cpu_cores ]]; then
         if [[ ! $parallelism_ci_cpu_cores =~ ^[[:digit:]]+$ ]]; then
@@ -233,21 +239,24 @@ function common::get_cpu_num {
       return
     fi
 
-    echo $((millicpu / 1000))
+    cpu_num=$((cpu_quota / cpu_period))
+    [[ $cpu_num -lt 1 ]] && echo 1 || echo $cpu_num
     return
   fi
 
   if [[ -f /sys/fs/cgroup/cpu.max ]]; then
     # Inside Linux (Docker?) container
-    millicpu=$(cut -d' ' -f1 /sys/fs/cgroup/cpu.max)
+    cpu_quota=$(cut -d' ' -f1 /sys/fs/cgroup/cpu.max)
+    cpu_period=$(cut -d' ' -f2 /sys/fs/cgroup/cpu.max)
 
-    if [[ $millicpu == max ]]; then
+    if [[ $cpu_quota == max || $cpu_period -lt 1 ]]; then
       # No limits
       nproc 2> /dev/null || echo 1
       return
     fi
 
-    echo $((millicpu / 1000))
+    cpu_num=$((cpu_quota / cpu_period))
+    [[ $cpu_num -lt 1 ]] && echo 1 || echo $cpu_num
     return
   fi