bug#21460: Race condition in tests/tail-2/assert.sh

["Pádraig Brady" <P@draigBrady.com>]

[-- Attachment #1: Type: text/plain, Size: 776 bytes --]

On 12/09/15 03:22, Pádraig Brady wrote:
> On 12/09/15 02:09, Paul Eggert wrote:
>> Pádraig Brady wrote:
>>> +  else if (new_stats.st_nlink == 0) /* XXX: what about multi-linked files.  */
>>
>> That comment was my thought exactly.  It appears to be a kernel bug that really 
>> needs to get fixed in the kernel; there just doesn't seem to be a reliable 
>> workaround in user space.
> 
> I double checked with kernel guys, and Al Viro
> essentially said the inode and directory operations are independent.
> https://lkml.org/lkml/2015/9/11/790
> 
> So we probably need to handle the IN_DELETE event
> on the directory to cater for this race, as done in:
> https://lists.gnu.org/archive/html/coreutils/2015-07/msg00015.html

I'll apply the attached later.

thanks,
Pádraig


[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: tail-unlink-notification-race.patch --]
[-- Type: text/x-patch; name="tail-unlink-notification-race.patch", Size: 3421 bytes --]

From 07ca8a227e04f9fb7bb0b21968056a562b8c2f83 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?P=C3=A1draig=20Brady?= <P@draigBrady.com>
Date: Thu, 2 Jul 2015 08:41:25 +0100
Subject: [PATCH] tail: handle kernel dentry unlink race
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Avoid the intermittent loss of "... has become inaccessible" messages.
That would cause tests/tail-2/assert.sh to fail sometimes,
mainly on uniprocessor systems.

* src/tail.c (tail_forever_inotify): Also monitor IN_DELETE
events on the directory, to avoid a dentry unlink()..open() race,
where the open() on the deleted file was seen to succeed after an,
unlink() and a subsequent IN_ATTRIB, was sent to tail.  Note an
IN_ATTRIB is sent on the monitored file to indicate the change in
number of links, and we can't just use a decrease in the number of
links to determine the file being unlinked, due to the possibility
of the file having multiple links.

Reported by Assaf Gordon and Ludovic Courtès.
Fixes http://bugs.gnu.org/21460
---
 src/tail.c | 21 ++++++++++++++-------
 1 file changed, 14 insertions(+), 7 deletions(-)

diff --git a/src/tail.c b/src/tail.c
index f916d74..7a34b0b 100644
--- a/src/tail.c
+++ b/src/tail.c
@@ -1429,8 +1429,8 @@ tail_forever_inotify (int wd, struct File_spec *f, size_t n_files,
                /* It's fine to add the same directory more than once.
                   In that case the same watch descriptor is returned.  */
               f[i].parent_wd = inotify_add_watch (wd, dirlen ? f[i].name : ".",
-                                                  (IN_CREATE | IN_MOVED_TO
-                                                   | IN_ATTRIB));
+                                                  (IN_CREATE | IN_DELETE
+                                                   | IN_MOVED_TO | IN_ATTRIB));
 
               f[i].name[dirlen] = prev;
 
@@ -1619,9 +1619,16 @@ tail_forever_inotify (int wd, struct File_spec *f, size_t n_files,
 
           fspec = &(f[j]);
 
-          /* Adding the same inode again will look up any existing wd.  */
-          int new_wd = inotify_add_watch (wd, f[j].name, inotify_wd_mask);
-          if (new_wd < 0)
+          int new_wd = -1;
+          bool deleting = !! (ev->mask & IN_DELETE);
+
+          if (! deleting)
+            {
+              /* Adding the same inode again will look up any existing wd.  */
+              new_wd = inotify_add_watch (wd, f[j].name, inotify_wd_mask);
+            }
+
+          if (! deleting && new_wd < 0)
             {
               if (errno == ENOSPC || errno == ENOMEM)
                 {
@@ -1641,7 +1648,7 @@ tail_forever_inotify (int wd, struct File_spec *f, size_t n_files,
           /* This will be false if only attributes of file change.  */
           bool new_watch = fspec->wd < 0 || new_wd != fspec->wd;
 
-          if (new_watch)
+          if (! deleting && new_watch)
             {
               if (0 <= fspec->wd)
                 {
@@ -1683,7 +1690,7 @@ tail_forever_inotify (int wd, struct File_spec *f, size_t n_files,
       if (! fspec)
         continue;
 
-      if (ev->mask & (IN_ATTRIB | IN_DELETE_SELF | IN_MOVE_SELF))
+      if (ev->mask & (IN_ATTRIB | IN_DELETE | IN_DELETE_SELF | IN_MOVE_SELF))
         {
           /* Note for IN_MOVE_SELF (the file we're watching has
              been clobbered via a rename) we leave the watch
-- 
2.5.0