syz-hub: which corpus to sync with the GCE one?

[Matthieu Baerts]

Hello,

On my side, I have 3 to 4 machines running syz-manager with different
kernel configs [1] and one machine is also running syz-hub. Initially, I
started from an empty corpus. I understood that it would be better for
me to re-use syzbot corpus [2]. How should I do that?

Do I simply need to stop all instances, replace syz-hub's corpus, and
relaunch everything? Or merge the corpuses? Or also replace the ones
linked to each syz-manager instances?

Also, are different corpuses available which might be more suited to my
case with instances having different kconfig (with/without KASAN, etc.
[1]) and focussed on the network area?

[1] https://github.com/multipath-tcp/mptcp_net-next/wiki/Testing#syzkaller
[2]
https://storage.googleapis.com/syzkaller/corpus/ci-upstream-kasan-gce-corpus.db

Cheers,
Matt
--
Sponsored by the NGI0 Core fund.

[Dmitry Vyukov]

Hi Matt,

IIRC the syz-hub corpus has slightly different format.
What I would probalby do is:
- stop all managers
- use tools/syz-db to merge the syzbot corpus into your existing corpus:
https://github.com/google/syzkaller/blob/master/tools/syz-db/syz-db.go#L101-L102
- start managers connected to syz-hub

It will take some time after first start for managers to chew the
large corpus, but at some potint they will re-minimize it and you will
get minimal corpus on each manager again.

Btw, I think at some point we changed syz-hub so that now it will send
programs to other managers only once per manager restart.
We had too high (unuseful) traffic on syzbot across instances, and
this helps to minimize the traffic. But it aims the syzbot/syz-ci
model when manager instances are periodically restarted.

[Matthieu Baerts]

Hi Dmitry,

Thank you for your reply!

On 16/02/2026 06:35, Dmitry Vyukov wrote:
> On Fri, 13 Feb 2026 at 19:51, 'Matthieu Baerts' via syzkaller
> <syzk...@googlegroups.com> wrote:
>>
>> Hello,
>>
>> On my side, I have 3 to 4 machines running syz-manager with different
>> kernel configs [1] and one machine is also running syz-hub. Initially, I
>> started from an empty corpus. I understood that it would be better for
>> me to re-use syzbot corpus [2]. How should I do that?
>>
>> Do I simply need to stop all instances, replace syz-hub's corpus, and
>> relaunch everything? Or merge the corpuses? Or also replace the ones
>> linked to each syz-manager instances?
>>
>> Also, are different corpuses available which might be more suited to my
>> case with instances having different kconfig (with/without KASAN, etc.
>> [1]) and focussed on the network area?
>>
>> [1] https://github.com/multipath-tcp/mptcp_net-next/wiki/Testing#syzkaller
>> [2]
>> https://storage.googleapis.com/syzkaller/corpus/ci-upstream-kasan-gce-corpus.db
>
> Hi Matt,
>
> IIRC the syz-hub corpus has slightly different format.
> What I would probalby do is:

Many thanks for these instructions!

> - stop all managers
> - use tools/syz-db to merge the syzbot corpus into your existing corpus:
> https://github.com/google/syzkaller/blob/master/tools/syz-db/syz-db.go#L101-L102

Just to be sure, which local corpus should I merge with the syzbot one?
The one linked to syz-hub, or the one linked to each syz-manager
instance? Or all of them?

> - start managers connected to syz-hub
>
> It will take some time after first start for managers to chew the
> large corpus, but at some potint they will re-minimize it and you will
> get minimal corpus on each manager again.

Great!

> Btw, I think at some point we changed syz-hub so that now it will send
> programs to other managers only once per manager restart.
> We had too high (unuseful) traffic on syzbot across instances, and
> this helps to minimize the traffic. But it aims the syzbot/syz-ci
> model when manager instances are periodically restarted.

Good to know, thank you! I usually restart each syz-manager once a week,
with an updated kernel and syzkaller suite. When all syz-manager
instances are off, I also update and restart syz-hub, then restart each
syz-manager.

[Dmitry Vyukov]

Each manager's corpus.

> > - start managers connected to syz-hub
> >
> > It will take some time after first start for managers to chew the
> > large corpus, but at some potint they will re-minimize it and you will
> > get minimal corpus on each manager again.
>
> Great!
>
> > Btw, I think at some point we changed syz-hub so that now it will send
> > programs to other managers only once per manager restart.
> > We had too high (unuseful) traffic on syzbot across instances, and
> > this helps to minimize the traffic. But it aims the syzbot/syz-ci
> > model when manager instances are periodically restarted.
>
> Good to know, thank you! I usually restart each syz-manager once a week,
> with an updated kernel and syzkaller suite. When all syz-manager
> instances are off, I also update and restart syz-hub, then restart each
> syz-manager.

You may try to use syz-ci to do periodic updates of both kernel and
syzkaller. You would need to configure kernel/syzkaller git
repo/branch, and it will do own builds of both.

[Matthieu Baerts]

Thank you, I will do that and *not* modify the syz-hub's one directly!

>>> - start managers connected to syz-hub
>>>
>>> It will take some time after first start for managers to chew the
>>> large corpus, but at some potint they will re-minimize it and you will
>>> get minimal corpus on each manager again.
>>
>> Great!
>>
>>> Btw, I think at some point we changed syz-hub so that now it will send
>>> programs to other managers only once per manager restart.
>>> We had too high (unuseful) traffic on syzbot across instances, and
>>> this helps to minimize the traffic. But it aims the syzbot/syz-ci
>>> model when manager instances are periodically restarted.
>>
>> Good to know, thank you! I usually restart each syz-manager once a week,
>> with an updated kernel and syzkaller suite. When all syz-manager
>> instances are off, I also update and restart syz-hub, then restart each
>> syz-manager.
>
> You may try to use syz-ci to do periodic updates of both kernel and
> syzkaller. You would need to configure kernel/syzkaller git
> repo/branch, and it will do own builds of both.

Thank you, I will look at syz-ci.

Is it OK to restart syz-manager each time a new commit is pushed? I
mean: will syzkaller be able to find new paths if it is restarted a few
times a day?

[Dmitry Vyukov]

syz-ci batches updates, and restarts managers at most every 24h IIRC.

[Matthieu Baerts]

Ah OK, it makes sense. I thought I had to let syz-manager running for a
few days to have a good coverage, but if 24h is enough, that's good!