Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 8 Jan 2021 20:17:25 +0300
From:      Vasily Postnicov <shamaz.mazum@gmail.com>
To:        Mark Johnston <markj@freebsd.org>
Cc:        freebsd-net@freebsd.org
Subject:   Re: DNS using Name Service Switch module and Casper
Message-ID:  <CADnZ6Bm96bjJN5gcpCWiNKbNou3XvxZmCD2-YbX34%2B00L=UdPw@mail.gmail.com>
In-Reply-To: <X/h%2BJRmXmrOfmXBM@raichu>
References:  <CADnZ6Bke=9%2B_pMc6rkbheNUWS-H6_X14%2Bf%2BWz5cfUCD=BTwk=g@mail.gmail.com> <X/R7Ahz8sz5v%2BoFa@raichu> <CADnZ6BmUJxVZx155j8opJKNsHJBE5mWz9D=MBE0Y_xu-kgOBfQ@mail.gmail.com> <X/h%2BJRmXmrOfmXBM@raichu>

next in thread | previous in thread | raw e-mail | index | archive | help
I have noticed that after I kill stuck ping, the process spawned with
cap_init() remains. I cannot even kill it with SIGKILL. This is the
output of procstat on such a process.


vasily   969   0.0  0.1   26428   6532 v0  I    22:43    0:00.00 ping
vonbraun.local
vasily   983   0.0  0.1   26428   6532 v0  I    22:43    0:00.00 ping
resurrected.local
vasily  1024   0.0  0.1   26428   6532 v0  I    22:49    0:00.00 ping
resurrected.local
vasily  1028   0.0  0.1   26428   6532 v0  I    22:49    0:00.00 ping
resurrected.local
root    1089   0.0  0.0   12976   2512 v1  S+   22:58    0:00.01 grep ping
  PID    TID COMM                TDNAME              KSTACK
 1028 100579 ping                -                   mi_switch+0x155
sleepq_switch+0x109 sleepq_catch_signals+0x266 sleepq_wait_sig+0x9
_sleep+0x2aa umtxq_sleep+0x19e do_lock_umutex+0x744
__umtx_op_wait_umutex+0x49 sys__umtx_op+0x7a amd64_syscall+0x12e
fast_syscall_common+0xf8

I checked ZeroMQ on which my NSS module is based. It does not use
pthread_atfork(), but uses lots of other unusual pthread functions,
like pthread_setaffinity_np() or pthread_setschedparam(). Do not know
if it matters. Also I do not quite understand when the code in my
module is executed. It should be executed after the capsicumized
sandbox is created, should it not? And I got hang in the process of
creating the sandbox. So I do not understand how my code affects this
process :)


=D0=BF=D1=82, 8 =D1=8F=D0=BD=D0=B2. 2021 =D0=B3. =D0=B2 18:45, Mark Johnsto=
n <markj@freebsd.org>:
>
> On Wed, Jan 06, 2021 at 07:08:14PM +0300, Vasily Postnicov wrote:
> > That's what I found.
> >
> > At first, ping calls cap_init() in capdns_setup(). cap_init() forks a
> > process, then the parent returns and the child calls casper_main_loop()=
.
> > The child and the parent both have a socket to communicate.
> > casper_main_loop() calls zygote_init() and that one blocks on fork(). I=
 do
> > not know how it could be. How can fork() block?
>
> Does you module somehow use pthread_atfork()?
>
> > The parent process later calls cap_service_open() and that function cal=
ls
> > cap_xfer_nvlist(). Because the child process is stuck somewhere in
> > zygote_init() it never sends an nvlist back. So ping blocks.
>
> Can you show output from "procstat -kk <pid>" when this hang occurs?
>
> > All this is figured out by inserting printf()'s. LLDB refuses to run pi=
ng
> > with 'error: Child exec failed'.
>
> Presumably it needs to be run as root since ping(8) is a setuid
> executable.
>
> > =D0=B2=D1=82, 5 =D1=8F=D0=BD=D0=B2. 2021 =D0=B3. =D0=B2 17:43, Mark Joh=
nston <markj@freebsd.org>:
> >
> > > On Tue, Jan 05, 2021 at 10:02:37AM +0300, Vasily Postnicov wrote:
> > > > Hello. I wrote a simple daemon called ZeroDNS which provides
> > > functionality
> > > > similar to multicast DNS, namely it discovers other participating
> > > machines
> > > > over the LAN and stores their hostname and IPv4 address pairs.
> > > >
> > > > Here is a NSS module which allows the system to use information fro=
m that
> > > > daemon:
> > > > https://github.com/shamazmazum/nss-zero-dns
> > > >
> > > > You need to modify /etc/nsswitch.conf, changing the line 'hosts: fi=
les
> > > dns'
> > > > to 'hosts: files dns zerodns'.
> > > >
> > > > It all works on FreeBSD 12.2-RELEASE, but sometimes not on 13.0-CUR=
RENT.
> > > > For example, ping(8) just blocks when trying to ping a host whose n=
ame is
> > > > resolvable with ZeroDNS. Turns out that programs built with casper
> > > support
> > > > (like ping(8) and some others) stop working with my NSS module (the=
y just
> > > > block trying to resolve the name).
> > >
> > > Presumably it's the casper process (i.e., cap_dns) that uses your
> > > module?  If the main ping process is blocked trying to resolve a name=
,
> > > it's waiting for the cap_dns process - where exactly is it getting
> > > stuck?
> > >
> > > > Is there some kind of manual on how to write casper-compatible NSS
> > > modules?
> > >



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CADnZ6Bm96bjJN5gcpCWiNKbNou3XvxZmCD2-YbX34%2B00L=UdPw>