Fix segmentation fault in vlog.Read

[jarifibrahim]

Fixes #1136 and #1131

The problem

In both the reported issues the segmentation fault would occur after vlog.DropAll (or if the log file was deleted). In case of dropAll, here's how we delete files
https://github.com/dgraph-io/badger/blob/8b99eb433aa799029b2e4cecc3c6e9b19c527490/value.go#L739-L757

At the same time, there could be a read on the same vlog file. Here's the code that reads from the file
https://github.com/dgraph-io/badger/blob/8b99eb433aa799029b2e4cecc3c6e9b19c527490/value.go#L1411-L1419

The problem occurs if we try to access the buf returned by readValueBytes. There is a window in which if we try to access the buf, we get a segmentation fault.

The following sequence of events would cause the segmentation fault.

1. vlog.Read is called and the control is at line 1411
   https://github.com/dgraph-io/badger/blob/8b99eb433aa799029b2e4cecc3c6e9b19c527490/value.go#L1411-L1419
2. vlog.DropAll is called at the same time. vlog.DropAll completes while the control is at line 1411 in vlog.Read. This essentially means all the log files are deleted and a new vlog0 file is created. This file has a fmap (mmap buffer) of 2 GB and the file size is 0 bytes (since we haven't written anything)
3. Now, the execution continues on vlog.Read and header.Decode (which is were the buf accessed for the first time) crashes with Segmentation fault ( SIGBUS error). As mentioned in comment Fix value log dropAll race condition #1140 (comment), the SIGBUS occurs when we try to access a mmap buffer which is beyond the actual file size.

With the following changes, I was able to reproduce the segmentation fault. Notice that we stop the execution at line 1411 in vlog.Read until DropAll completes.

diff --git a/managed_db_test.go b/managed_db_test.go
index 5a25174..c4e27ba 100644
--- a/managed_db_test.go
+++ b/managed_db_test.go
@@ -209,7 +209,10 @@ func TestDropAllWithPendingTxn(t *testing.T) {
 	wg.Add(1)
 	go func() {
 		defer wg.Done()
-		itr := txn.NewIterator(DefaultIteratorOptions)
+		d := DefaultIteratorOptions
+		// This is necessary otherwise we'll run into a deadlock in this test.
+		d.PrefetchValues = false
+		itr := txn.NewIterator(d)
 		defer itr.Close()
 
 		var keys []string
@@ -246,6 +249,7 @@ func TestDropAllWithPendingTxn(t *testing.T) {
 		defer wg.Done()
 		time.Sleep(2 * time.Second)
 		require.NoError(t, db.DropAll())
+		db.vlog.ch <- struct{}{}
 	}()
 	wg.Wait()
 }
diff --git a/value.go b/value.go
index b28b2b1..a63a519 100644
--- a/value.go
+++ b/value.go
@@ -780,6 +780,7 @@ type lfDiscardStats struct {
 
 type valueLog struct {
 	dirPath string
+	ch      chan struct{}
 	elog    trace.EventLog
 
 	// guards our view of which files exist, which to be deleted, how many active iterators
@@ -1004,6 +1005,7 @@ func (vlog *valueLog) replayLog(lf *logFile, offset uint32, replayFn logEntry) e
 func (vlog *valueLog) open(db *DB, ptr valuePointer, replayFn logEntry) error {
 	vlog.opt = db.opt
 	vlog.db = db
+	vlog.ch = make(chan struct{}, 1)
 	// We don't need to open any vlog files or collect stats for GC if DB is opened
 	// in InMemory mode. InMemory mode doesn't create any files/directories on disk.
 	if vlog.opt.InMemory {
@@ -1413,6 +1415,7 @@ func (vlog *valueLog) Read(vp valuePointer, s *y.Slice) ([]byte, func(), error)
 			"Invalid value pointer offset: %d greater than current offset: %d",
 			vp.Offset, vlog.woffset())
 	}
+	<-vlog.ch
 	buf, lf, err := vlog.readValueBytes(vp, s)
 	// log file is locked so, decide whether to lock immediately or let the caller to
 	// unlock it, after caller uses it.

---

This change is 

[coveralls]

Coverage decreased (-7.9%) to 69.891% when pulling 5ccc431 on ibrahim/header-decode-fix into 8b99eb4 on master.

[manishrjain]

Reviewed 1 of 1 files at r1.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on @ashish-goswami)