Re: tail is reading already rotated file.

[최종우]

To reproduce the problem, you can run this script. It exhausted inotify watches. (in order to get the result more quickly, decrease the value of max_user_watches.)

let count=0; echo $count>>a.log; ./tail -F a.log & while true; do mv a.log a.log.$count; let count=count+1; echo $count>>a.log; done

I tested this problem in coreutils 8.4 and 8.23.

2015-02-03 20:16 GMT+09:00 Pádraig Brady <address@hidden>:

On 03/02/15 07:26, 최종우 wrote:
> tail doesn't release any watch descriptors watching already rotated files except for file deletion. Does it?
>
> In my case, the number of opened watch descriptors exceeded a value of /proc/sys/fs/inotify/max_user_watches and tail was opening a renamed file.
> I've tested the following steps.
>
> 1) cat /proc/sys/fs/inotify/max_user_watches
>
> 2) log files are generated and tail -F the latest log file.
>
> (The number of log files exceeds max_user_watches and tail doesn't print any output to stdout.)
>
> 3) ls -l /proc/pid/fd
>
> (tailing already rotated file. It seems that tail couldn't keep track of the lastest file because of a shortage of watch descriptors.)
>
>
> And I think that the latest revison of tail has the same issue. May I ask you the reason why you don't release watch descriptors for renamed files?

I don't see where we leak watch descriptors, and if we run out of resources
we should revert to polling mode. Can you reproduce the problem with
tail built from the coreutils 8.23?

  wget ftp://ftp.gnu.org/pub/gnu/coreutils/coreutils-8.23.tar.gz
  tar -xf coreutils-8.23.tar.gz
  cd coreutils-8.23
  ./configure --quiet && make

you can then use src/tail for testing.

What are the exact steps to reproduce if 8.23 has the issue?

thanks,
Pádraig.