duplicate devices

[stintel]

Background information

• Dendrite version or git SHA: 0.9.6
• Monolith or Polylith?: monolith
• SQLite3 or Postgres?: postgres
• Running in Docker?: no
• go version: 1.18.5
• Client used (if applicable): matrix-ircd cdf5fc70471d21c8d631e98ef19e04de83f27390

Description

• What is the problem: duplicate devices for same client
• Who is affected: clients connected via matrix-ircd on my server
• How is this bug manifesting: when using matrix-ircd, a lot of duplicate devices are appearing, sometimes every 5 minutes
• When did this first appear: for as long as I can remember, started using dendrite at version 0.3.4 in December 2020, matrix-ircd at commit hash 9f7794b6ff7a661e5a45882ef4c29274990814c4 also in December 2020.

Steps to reproduce

• setup dendrite
• setup matrix-ircd
• connect an irc client to matrix-ircd
• wait for irc client to disconnect due to Dendrite sending invalid packets to matrix-ircd (e.g. Unhandled IO error, error: Sync returned invalid JSON: missing field join at line 1 column 268)
• wait for irc client to reconnect - each attempt creates a new device

While possibly a neochat bug, I belive Dendrite should have a configurable limit for max # of devices per user to avoid situations like this. The default should be reasonably low. I have just signed out >600 (!) devices, both dendrite and postgresql where eating all 4 vCPUs after doing so, with load average growing over 10.

I believe this could be related to my Dendrite instance permanently eating a CPU core and consuming consistently more than 1GB of RAM, and random rooms being completely out of sync for the last few months. Possibly also the cause of #2566? I have had that problem many times, and deleting the jetstream directory seemed to bring temporary relief, but eventually some channels stopped syncing again.

I've experienced issues for a while, but what finally led me to start digging was various room names being missing, instead member names separated by comma being displayed. Unfortunately I forgot to screenshot that and that problem seems to have solved after signing out all the duplicate devices.

[stintel]

Not so long after opening this issue the dendrite-monolith-server process is back to eating a CPU core and > 1GB RAM, but at least clients seem to respond normally now, seeing proper room names and no longer crazy slow.

I wanted to ask this in the Dendrite room, but it appears my server is banned from the room (already in contact with abuse@ about that), so I'll ask it here: Do the profile files generated from the instructions on https://matrix-org.github.io/dendrite/development/profiling contain sensitive data?

[neilalexander]

The profiles are safe, so might be worth inspecting what's going on there.

Has this bug been reported to the NeoChat team? Something is horribly wrong if it's registering new devices every few minutes.

[stintel]

Has this bug been reported to the NeoChat team? Something is horribly wrong if it's registering new devices every few minutes.

Actually the client causing this might not have been NeoChat. It appears to be matrix-ircd, which for some reason shows up as a different client. After enabling real_ip_header I now see the correct client IP, which confirms it's matrix-ircd and not NeoChat. I have updated the issue to reflect this.

Looks like this is a known issue in matrix-ircd: https://github.com/matrix-org/matrix-ircd/issues/43. Nonetheless, I believe the possibility to limit the amount of devices a user can register is a reasonable feature request?

[kegsay]

I believe the possibility to limit the amount of devices a user can register is a reasonable feature request?

YES PLEASE. The problem is that if you omit the device_id field on registration/login, one is randomly generated. So just innocently missing this field can cause rampant device creation. This then exacerbates slow E2EE because there's so many devices to encrypt for.