linux-ext4 - Realtime behavior of named pipes

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [day] [month] [year] [list]

Message-ID: <bca722097ded477dbf62d5896bf01656@THSUMIDEUMSG03P.deu.umi.grp>
Date:   Thu, 31 Mar 2022 08:53:22 +0000
From:   HARDERS Bernd <Bernd.Harders@...lesgroup.com>
To:     "linux-ext4@...r.kernel.org" <linux-ext4@...r.kernel.org>
CC:     PFLANZ Roger <Roger.Pflanz@...lesgroup.com>
Subject: Realtime behavior of named pipes

Hello, 

my name is Bernd Harders, I'm working for the thales company and I just have analyzed a real-time issue in context of named pipes, ext4 and jbd2, using ftrace and trace compass.

While the issue can be fixed by moving the pipe to the ramfs, I want to understand what's going on under the hood. 
I'm also interested in the question, if this issue belongs to the use of the very old RedHat7 and kernel 3.10, or if this behavior is still to be expectable. 

Here comes a short description of the behavior:

We have two high priority realtime threads, which communicate and synchronize in a high frequent manner using a named pipe. The named pipe directory entry is located on an ext4 partition, a lvm is in use. 

I found that in seldom cases the threads are blocked  for milliseconds, while a block_io, forced by the jbd2/dm thread is active. In this case there is no deterministic latency anymore, because it seems that the duration of the blocking also depends on other block_io's.  

Our idea is that this behavior is related to timestamp updates in the directory, forcing the journaling activities. 

While I can totally understand that io to a file is blocked, when journaling is doing its work, it is not obviously clear why this also happen on the pipe, because as I'm told the pipe io internally works to memory. We only write and read an integer value per transfer, so there is no need for temporary file io storage.

I'm looking forward to any explanations.

Best regards
Bernd Harders

{OPEN}

Bernd Harders
Senior Software Engineer
Naval Systems
Thales Deutschland
Switchboard Phone: +49 431-88737-344
Mobile: +49 172 828 14 50

E-Mail:   bernd.harders@...lesgroup.com

Thales Deutschland GmbH
Am Kiel-Kanal 3 - 24106 Kiel - Germany

[DEU] Achtung: Neue Anschrift! 
[ENG] Attention: New address!
www.thalesgroup.com

Sitz der Gesellschaft/Domicile of the Company: Stuttgart
Amtsgericht/District Court: Stuttgart HRB 728793
Geschäftsführer/Managing Directors:
Oliver Dörre (Vorsitzender/Chairman), Dr. Henning Biebinger, Dirk J.H. de Bruijn, Dr. Yves Joannic
Aufsichtsratsvorsitzender/Chairman of the Supervisory Board: Bernhard Gerwert