diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go
index 70578d3fa..335003dd8 100644
--- a/pkg/sentry/fs/proc/task.go
+++ b/pkg/sentry/fs/proc/task.go
@@ -461,7 +461,14 @@ func (s *taskStatData) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle)
 	fmt.Fprintf(&buf, "%d %d ", linux.ClockTFromDuration(cputime.UserTime), linux.ClockTFromDuration(cputime.SysTime))
 	fmt.Fprintf(&buf, "%d %d ", s.t.Priority(), s.t.Niceness())
 	fmt.Fprintf(&buf, "%d ", s.t.ThreadGroup().Count())
-	fmt.Fprintf(&buf, "0 0 " /* itrealvalue starttime */)
+
+	// itrealvalue. Since kernel 2.6.17, this field is no longer
+	// maintained, and is hard coded as 0.
+	fmt.Fprintf(&buf, "0 ")
+
+	// Start time is relative to boot time, expressed in clock ticks.
+	fmt.Fprintf(&buf, "%d ", linux.ClockTFromDuration(s.t.StartTime().Sub(s.t.Kernel().Timekeeper().BootTime())))
+
 	var vss, rss uint64
 	s.t.WithMuLocked(func(t *kernel.Task) {
 		if mm := t.MemoryManager(); mm != nil {
diff --git a/test/syscalls/linux/proc.cc b/test/syscalls/linux/proc.cc
index 0da682e7b..64d75d9fb 100644
--- a/test/syscalls/linux/proc.cc
+++ b/test/syscalls/linux/proc.cc
@@ -880,6 +880,14 @@ TEST_P(ProcPidStatTest, HasBasicFields) {
   EXPECT_EQ("R", fields[2]);  // task state
   EXPECT_EQ(absl::StrCat(getppid()), fields[3]);
 
+  // If the test starts up quickly, then the process start time and the kernel
+  // boot time will be very close, and the proc starttime field (which is the
+  // delta of the two times) will be 0.  For that unfortunate reason, we can
+  // only check that starttime >= 0, and not that it is strictly > 0.
+  uint64_t starttime;
+  ASSERT_TRUE(absl::SimpleAtoi(fields[21], &starttime));
+  EXPECT_GE(starttime, 0);
+
   uint64_t vss;
   ASSERT_TRUE(absl::SimpleAtoi(fields[22], &vss));
   EXPECT_GT(vss, 0);