Git designed for AI agents.
package cgroups
import (
"io"
"os/exec"
"github.com/feder-cr/git-ai/internal/dontpanic"
"github.com/feder-cr/git-ai/internal/gitaly/config/cgroups"
"github.com/feder-cr/git-ai/internal/gitaly/storage"
"github.com/feder-cr/git-ai/internal/log"
)
type addCommandCfg struct {
cgroupKey string
}
// PSIData holds the pressure stall information averages and total.
// These metrics indicate the percentage of wall-clock time tasks were stalled.
type PSIData struct {
Avg10 float64 // 10-second average percentage
Avg60 float64 // 60-second average percentage
Avg300 float64 // 300-second (5 minute) average percentage
Total uint64 // Cumulative total stall time in microseconds
}
// PSIMetrics holds both "some" and "full" pressure metrics.
// "Some" means at least one task was stalled; "Full" means all tasks were stalled.
type PSIMetrics struct {
Some PSIData
Full PSIData
}
// CgroupStats stores the current usage statistics of the resources managed by
// cgroup manager. They are fetched from the cgroupfs statistic files.
type CgroupStats struct {
// Path is the path to the cgroup
Path string
// CPUThrottledCount is much the CPU was throttled, this field is fetched from the `nr_throttled` field of
// `cpu.stat` file.
CPUThrottledCount uint64
// CPUThrottledCount is the accumulated duration of CPU throttled time in seconds. It is from the
// `throttled_time` (V1) or the `throttled_usec` (v2) field of the `cpu.stat` file.
CPUThrottledDuration float64
// MemoryUsage is the current memory usage in bytes of a cgroup. It's fetched from the `memory.usage_in_bytes`
// (V1) or the `memory.current` (V2) files.
MemoryUsage uint64
// MemoryLimit is the current memory limit in bytes of a cgroup. It's from the `memory.limit_in_bytes` (V1) or
// the `memory.max` (V2) files.
MemoryLimit uint64
// OOMKills is the accumulated count when the cgroup is being OOM-ed. It's read from the `memory.oom_control`
// (V1) and the `memory.events` (V2) files.
OOMKills uint64
// UnderOOM is the current OOM status of a cgroup. This information is available for Cgroup V1 only. It's read
// from the `memory.oom_control` file.
UnderOOM bool
// TotalAnon reflects the `rss` (V1) or `anon` (V2) of `memory.stat` file. That's amount of memory used in anonymous mappings
TotalAnon uint64
// ActiveAnon reflects the `active_file` of `memory.stat` file, anonymous and swap cache memory on active LRU list.
TotalActiveAnon uint64
// InactiveAnon reflects the `inactive_anon` of `memory.stat` file, anonymous and swap cache memory on inactive LRU list.
TotalInactiveAnon uint64
// TotalFile reflects the `cache` (V1) or `file` (V2) of `memory.stat` file, bytes of page cache memory.
TotalFile uint64
// TotalActiveFile reflects the `active_file` of `memory.stat` file, bytes of file-backed memory that cannot be reclaimed.
TotalActiveFile uint64
// TotalInactiveFile reflects the `inactive_file` of `memory.stat` file, bytes of file-backed memory on inactive LRU list.
TotalInactiveFile uint64
// PSI Memory Pressure from `memory.pressure` file (cgroups v2 only).
MemoryPSI PSIMetrics
// PSI IO Pressure from `io.pressure` file (cgroups v2 only).
IOPSI PSIMetrics
// PSI CPU Pressure from `cpu.pressure` (cgroups v2 only, "Some" only).
CPUPSI PSIMetrics
// Memory Events from `memory.events` file (cgroups v2 only).
// These are counters that increment each time certain memory events occur.
// MemoryHighEvents is the number of times the cgroup's memory usage exceeded memory.high.
// A rising count indicates the cgroup is being throttled.
MemoryHighEvents uint64
// MemoryMaxEvents is the number of times the cgroup's memory usage was about to exceed memory.max.
// This indicates how close we are to triggering OOM.
MemoryMaxEvents uint64
// MemoryOOMEvents is the number of times the OOM killer was invoked in this cgroup.
// Note: This may differ from OOMKills which counts actual kills.
MemoryOOMEvents uint64
// PgFault is the total number of page faults incurred.
PgFault uint64
// PgMajFault is the number of major page faults incurred (required disk IO).
// High values indicate cache misses causing disk reads - a key indicator of cache pressure.
PgMajFault uint64
}
// Stats stores statistics of all cgroups managed by a manager
type Stats struct {
// ParentStats stores the statistics of the parent cgroup.
ParentStats CgroupStats
// isEmpty is an internal field used to identify if the Stats are from an enabled
// cgroup manager, or from a NoopManager. When set to true, all Stats values should
// be discarded (all values should be zero/empty values anyway) because it means the
// cgroup manager is not running. This is currently only used from the NoopManager.
isEmpty bool
// version holds the group version for this Stats instance.
// 0: Noop (not set)
// 1: v1
// 2: v2
version uint8
}
// IsEmpty returns true if the Stats fields are all set to
// zero/empty values. This is most likely because they are
// returned from the NoopManager.
func (s Stats) IsEmpty() bool {
return s.isEmpty
}
// CgroupV1 returns true of the cgroup version is 1
func (s Stats) CgroupV1() bool {
return !s.IsEmpty() && s.version == 1
}
// CgroupV2 returns true of the cgroup version is 2
func (s Stats) CgroupV2() bool {
return !s.IsEmpty() && s.version == 2
}
// AddCommandOption is an option that can be passed to AddCommand.
type AddCommandOption func(*addCommandCfg)
// WithCgroupKey overrides the key used to derive the Cgroup bucket. If not passed, then the command
// arguments will be used as the cgroup key.
func WithCgroupKey(cgroupKey string) AddCommandOption {
return func(cfg *addCommandCfg) {
cfg.cgroupKey = cgroupKey
}
}
// RepositoryKey returns the cgroup key for commands that operate on the given repository. All
// processes spawned on behalf of a repository should use this key so that they share its cgroup.
func RepositoryKey(repo storage.Repository) string {
return repo.GetStorageName() + "/" + repo.GetRelativePath()
}
// Manager supplies an interface for interacting with cgroups
type Manager interface {
// Setup creates cgroups and assigns configured limitations.
// It is expected to be called once at Gitaly startup from any
// instance of the Manager.
Setup() error
// Ready returns true if this cgroup manager is configured and
// ready to use.
Ready() bool
// AddCommand adds a Cmd to a cgroup.
AddCommand(*exec.Cmd, ...AddCommandOption) (string, error)
// SupportsCloneIntoCgroup returns whether this Manager supports the CloneIntoCgroup method.
// CloneIntoCgroup requires CLONE_INTO_CGROUP which is only supported with cgroups version 2
// with Linux 5.7 or newer.
SupportsCloneIntoCgroup() bool
// CloneIntoCgroup configures the cgroup parameters UseCgroupFD and CgroupFD in SysProcAttr
// to start the command directly in the correct cgroup. On success, the function returns an io.Closer
// that must be closed after the command has been started to close the cgroup's file descriptor.
CloneIntoCgroup(*exec.Cmd, ...AddCommandOption) (string, io.Closer, error)
// Stats returns cgroup accounting statistics collected by reading
// cgroupfs files. Those statistics are generic for both Cgroup V1
// and Cgroup V2.
Stats() (Stats, error)
}
// NewManager returns the appropriate Cgroups manager
func NewManager(cfg cgroups.Config, logger log.Logger, pid int) Manager {
if cfg.Repositories.Count < 1 {
if cfg.CPUShares > 0 || cfg.MemoryBytes > 0 {
logger.Warn("cgroups repositories.count value was not configured or set to 0. Cgroup support will be disabled. Please set a positive value for repositories.count")
}
return &NoopManager{}
}
return newCgroupManager(cfg, logger, pid)
}
// StartPruningOldCgroups prunes old cgroups for both the memory and cpu subsystems
// in a goroutine.
func StartPruningOldCgroups(cfg cgroups.Config, logger log.Logger) {
dontpanic.Go(logger, func() {
pruneOldCgroups(cfg, logger)
})
}