tail: avoid a race where we could miss new data with --pid

* src/tail.c (tail_forever, tail_forever_inotify): Close a race in
tail_forever_inotify where new data written after the file check by
a now dead process, but before the pid check, is not output.  We use
the POSIX guarantee that read() and write() are serialized wrt each
other even in separate processes, to assume full file consistency
after exit() and so poll for new data _after_ the writer has exited.
This also allows us to not redundantly _wait_ for new data if the
process is dead.
* tests/tail-2/pid: Remove the now partially invalid sub second sleep
check as we now don't unconditionally wait, and replace it with a check
for the redundant sleep.  Also clarify some of the existing comments.
* NEWS: Mention the fix.

3 files changed, 38 insertions, 29 deletions

diff --git a/NEWS b/NEWS
index e7e4ddccf..aff0744a1 100644
--- a/NEWS
+++ b/NEWS
@@ -22,6 +22,12 @@ GNU coreutils NEWS                                    -*- outline -*-
   from a failed stat/lstat.  For example ls -Lis now prints "?", not "0",
   for the inode number and allocated size of a dereferenced dangling symlink.
 
+  tail --follow --pid now avoids a race condition where data written
+  just before the process dies might not have been output by tail.
+  Also, tail no longer delays at all when the specified pid is not live.
+  [The race was introduced in coreutils-7.5,
+   and the unnecessary delay was present since textutils-1.22o]
+
 ** Portability
 
   On Solaris 9, many commands would mistakenly treat file/ the same as
diff --git a/src/tail.c b/src/tail.c
index f9f8c3904..f3fe6a341 100644
--- a/src/tail.c
+++ b/src/tail.c
@@ -1135,9 +1135,6 @@ tail_forever (struct File_spec *f, size_t n_files, double sleep_interval)
           if (writer_is_dead)
             break;
 
-          if (xnanosleep (sleep_interval))
-            error (EXIT_FAILURE, errno, _("cannot read realtime clock"));
-
           /* Once the writer is dead, read the files once more to
              avoid a race condition.  */
           writer_is_dead = (pid != 0
@@ -1146,6 +1143,10 @@ tail_forever (struct File_spec *f, size_t n_files, double sleep_interval)
                                signal to the writer, so kill fails and sets
                                errno to EPERM.  */
                             && errno != EPERM);
+
+          if (!writer_is_dead && xnanosleep (sleep_interval))
+            error (EXIT_FAILURE, errno, _("cannot read realtime clock"));
+
         }
     }
 }
@@ -1179,6 +1180,7 @@ tail_forever_inotify (int wd, struct File_spec *f, size_t n_files,
   Hash_table *wd_table;
 
   bool found_watchable = false;
+  bool writer_is_dead = false;
   int prev_wd;
   size_t evlen = 0;
   char *evbuf;
@@ -1266,30 +1268,30 @@ tail_forever_inotify (int wd, struct File_spec *f, size_t n_files,
          indefinetely.  */
       if (pid)
         {
-          fd_set rfd;
-          struct timeval select_timeout;
-          int n_descriptors;
-
-          FD_ZERO (&rfd);
-          FD_SET (wd, &rfd);
+          if (writer_is_dead)
+            exit (EXIT_SUCCESS);
 
-          select_timeout.tv_sec = (time_t) sleep_interval;
-          select_timeout.tv_usec = 1000000 * (sleep_interval
-                                              - select_timeout.tv_sec);
+          writer_is_dead = (kill (pid, 0) != 0 && errno != EPERM);
 
-          n_descriptors = select (wd + 1, &rfd, NULL, NULL, &select_timeout);
+          struct timeval delay; /* how long to wait for file changes.  */
+          if (writer_is_dead)
+            delay.tv_sec = delay.tv_usec = 0;
+          else
+            {
+              delay.tv_sec = (time_t) sleep_interval;
+              delay.tv_usec = 1000000 * (sleep_interval - delay.tv_sec);
+            }
 
-          if (n_descriptors == -1)
-            error (EXIT_FAILURE, errno, _("error monitoring inotify event"));
+           fd_set rfd;
+           FD_ZERO (&rfd);
+           FD_SET (wd, &rfd);
 
-          if (n_descriptors == 0)
-            {
-              /* See if the process we are monitoring is still alive.  */
-              if (kill (pid, 0) != 0 && errno != EPERM)
-                exit (EXIT_SUCCESS);
+           int file_change = select (wd + 1, &rfd, NULL, NULL, &delay);
 
-              continue;
-            }
+           if (file_change == 0)
+             continue;
+           else if (file_change == -1)
+             error (EXIT_FAILURE, errno, _("error monitoring inotify event"));
         }
 
       if (len <= evbuf_off)
diff --git a/tests/tail-2/pid b/tests/tail-2/pid
index 90f168430..760e289bc 100755
--- a/tests/tail-2/pid
+++ b/tests/tail-2/pid
@@ -29,26 +29,27 @@ touch here || framework_failure
 fail=0
 
 for inotify in ---disable-inotify ''; do
-  # Use tail itself to create a background process to monitor.
+  # Use tail itself to create a background process to monitor,
+  # which will auto exit when "here" is removed.
   tail -f $inotify here &
   bg_pid=$!
 
   # Ensure that tail --pid=PID does not exit when PID is alive.
-  timeout 1 tail -s.1 -f $inotify here --pid=$bg_pid
+  timeout 1 tail -f -s.1 --pid=$bg_pid $inotify here
   test $? = 124 || fail=1
 
   # Cleanup background process
   kill $bg_pid
 
-  # Ensure that tail --pid=PID exits successfully when PID is dead.
+  # Ensure that tail --pid=PID exits with success status when PID is dead.
   # Use an unlikely-to-be-live PID
-  timeout 3 tail -s.1 --pid=$PID_T_MAX -f $inotify /dev/null
+  timeout 3 tail -f -s.1 --pid=$PID_T_MAX $inotify /dev/null
   ret=$?
-  test $ret = 124 && skip_test_ "pid $PID_T_MAX present"
+  test $ret = 124 && skip_test_ "pid $PID_T_MAX present or tail too slow"
   test $ret = 0 || fail=1
 
-  # Ensure fractional sleep parameter is honored with --pid
-  timeout 3 tail -s.1 -f $inotify /dev/null --pid=$PID_T_MAX
+  # Ensure tail doesn't wait for data when PID is dead
+  timeout 3 tail -f -s10 --pid=$PID_T_MAX $inotify /dev/null
   test $? = 124 && fail=1
 done