Files
podman/pkg/api/server/idle/tracker.go
Alex Schultz 0ac857fc0f Change http ConnState actions between new and active
Currently it double counts connections because it's incrementing the
total for both the new and active states. Based on the comments, we
should only count new connections for the total count and perform the
timer stop actions when the connection has transitioned to an active
state.

Closes #8208

Signed-off-by: Alex Schultz <aschultz@redhat.com>
2020-10-31 20:31:33 -06:00

102 lines
3.4 KiB
Go

package idle
import (
"net"
"net/http"
"sync"
"time"
"github.com/sirupsen/logrus"
)
// Tracker holds the state for the server's idle tracking
type Tracker struct {
// Duration is the API idle window
Duration time.Duration
hijacked int // count of active connections managed by handlers
managed map[net.Conn]struct{} // set of active connections managed by http package
mux sync.Mutex // protect managed map
timer *time.Timer
total int // total number of connections made to this server instance
}
// NewTracker creates and initializes a new Tracker object
// For best behavior, duration should be 2x http idle connection timeout
func NewTracker(idle time.Duration) *Tracker {
return &Tracker{
managed: make(map[net.Conn]struct{}),
Duration: idle,
timer: time.NewTimer(idle),
}
}
// ConnState is called on HTTP connection state changes.
// - Once StateHijacked, StateClose is _NOT_ called on that connection
// - There are two "idle" timeouts, the http idle connection (not to be confused with the TCP/IP idle socket timeout)
// and the API idle window. The caller should set the http idle timeout to 2x the time provided to NewTacker() which
// is the API idle window.
func (t *Tracker) ConnState(conn net.Conn, state http.ConnState) {
t.mux.Lock()
defer t.mux.Unlock()
logrus.Debugf("IdleTracker %p:%v %dm+%dh/%dt connection(s)", conn, state, len(t.managed), t.hijacked, t.TotalConnections())
switch state {
case http.StateNew:
t.total++
case http.StateActive:
// stop the API timer when the server transitions any connection to an "active" state
t.managed[conn] = struct{}{}
t.timer.Stop()
case http.StateHijacked:
// hijacked connections should call Close() when finished.
// Note: If a handler hijack's a connection and then doesn't Close() it,
// the API timer will not fire and the server will _NOT_ timeout.
delete(t.managed, conn)
t.hijacked++
case http.StateIdle:
// When any connection goes into the http idle state, we know:
// - we have an active connection
// - the API timer should not be counting down (See case StateNew/StateActive)
break
case http.StateClosed:
oldActive := t.ActiveConnections()
// Either the server or a hijacking handler has closed the http connection to a client
if conn == nil {
t.hijacked-- // guarded by t.mux above
} else {
if _, found := t.managed[conn]; found {
delete(t.managed, conn)
} else {
logrus.Warnf("IdleTracker %p: StateClosed transition by un-managed connection", conn)
}
}
// Transitioned from any "active" connection to no connections
if oldActive > 0 && t.ActiveConnections() == 0 {
t.timer.Stop() // See library source for Reset() issues and why they are not fixed
t.timer.Reset(t.Duration) // Restart the API window timer
}
}
}
// Close is used to update Tracker that a StateHijacked connection has been closed by handler (StateClosed)
func (t *Tracker) Close() {
t.ConnState(nil, http.StateClosed)
}
// ActiveConnections returns the number of current managed or StateHijacked connections
func (t *Tracker) ActiveConnections() int {
return len(t.managed) + t.hijacked
}
// TotalConnections returns total number of connections made to this instance of the service
func (t *Tracker) TotalConnections() int {
return t.total
}
// Done is called when idle timer has expired
func (t *Tracker) Done() <-chan time.Time {
return t.timer.C
}