frostfs-node/pkg/local_object_storage/writecache/mode.go
Pavel Karpy e9461686b8 [#274] wc: Resolve possible deadlock
If operation with WC are _fast enough_ (e.g. `Init` failed and `Close` is
called immediately) there is a race and a deadlock that do not allow finish
(and start, in fact) an initialization routine because of taken `modeMtx`
and also do not allow finish `Close` call because of awaiting initialization
finish. So do stop initialization _before_ any mutex is taken.

Signed-off-by: Pavel Karpy <p.karpy@yadro.com>
2023-04-20 19:22:54 +03:00

94 lines
2.4 KiB
Go

package writecache
import (
"context"
"fmt"
"time"
"git.frostfs.info/TrueCloudLab/frostfs-api-go/v2/pkg/tracing"
"git.frostfs.info/TrueCloudLab/frostfs-node/internal/logs"
"git.frostfs.info/TrueCloudLab/frostfs-node/pkg/local_object_storage/shard/mode"
"git.frostfs.info/TrueCloudLab/frostfs-node/pkg/local_object_storage/util/logicerr"
"go.opentelemetry.io/otel/attribute"
"go.opentelemetry.io/otel/trace"
)
// ErrReadOnly is returned when Put/Write is performed in a read-only mode.
var ErrReadOnly = logicerr.New("write-cache is in read-only mode")
// ErrNotInitialized is returned when write-cache is initializing.
var ErrNotInitialized = logicerr.New("write-cache is not initialized yet")
// SetMode sets write-cache mode of operation.
// When shard is put in read-only mode all objects in memory are flushed to disk
// and all background jobs are suspended.
func (c *cache) SetMode(m mode.Mode) error {
ctx, span := tracing.StartSpanFromContext(context.TODO(), "writecache.SetMode",
trace.WithAttributes(
attribute.String("mode", m.String()),
))
defer span.End()
return c.setMode(ctx, m)
}
// setMode applies new mode. Must be called with cache.modeMtx lock taken.
func (c *cache) setMode(ctx context.Context, m mode.Mode) error {
var err error
turnOffMeta := m.NoMetabase()
if !c.initialized.Load() {
close(c.stopInitCh)
c.initWG.Wait()
c.stopInitCh = make(chan struct{})
defer func() {
if err == nil && !turnOffMeta {
c.initFlushMarks(ctx)
}
}()
}
c.modeMtx.Lock()
defer c.modeMtx.Unlock()
if turnOffMeta && !c.mode.NoMetabase() {
err = c.flush(ctx, true)
if err != nil {
return err
}
}
if c.db != nil {
if err = c.db.Close(); err != nil {
return fmt.Errorf("can't close write-cache database: %w", err)
}
}
// Suspend producers to ensure there are channel send operations in fly.
// flushCh is populated by `flush` with `modeMtx` taken, thus waiting until it is empty
// guarantees that there are no in-fly operations.
for len(c.flushCh) != 0 {
c.log.Info(logs.WritecacheWaitingForChannelsToFlush)
time.Sleep(time.Second)
}
if turnOffMeta {
c.mode = m
return nil
}
if err = c.openStore(m.ReadOnly()); err != nil {
return err
}
c.mode = m
return nil
}
// readOnly returns true if current mode is read-only.
// `c.modeMtx` must be taken.
func (c *cache) readOnly() bool {
return c.mode.ReadOnly()
}