[zeromq-dev] assertion at mailbox.cpp:182

[Dhammika Pathirana]

Hi Pavel,

On Sun, Jan 30, 2011 at 6:35 AM, Pavel Gushcha <pavimus at gmail.com> wrote:
> Hi, All!
>
> It seems, that I have the same problem as Zoufal Andreas:
> http://lists.zeromq.org/pipermail/zeromq-dev/2011-January/008914.html
> From some timepoint 2 applications from my system start to crash
> regularly with assert zmq_assert (new_sndbuf > old_sndbuf). Earlier i
> seen this assert sometimes while programs are finishing it work. But
> now, i see this error regularly (sometimes applications works 12
> hours, sometimes only 3).
>
> Info about my scenario:
> application1 has ST_PUSH TCP socket with setted ZMQ_HWM value to 10.
> I'm running 3 application instances on 3 different servers.
> application2 has ST_PULL TCP socket with setted ZMQ_HWM value to 1000.
> I'm running 1 application instance on 4-th server.
>
> Last time i seen following: first instance of application1 asserted,
> after some minutes asserted second instance of application1, after
> some minutes - asserts application2.
>
> I'm using zeromq-2.1.0 on hardened gentoo (kernel 2.6.36)
>
> There is my limits:
> # sysctl -a | grep net.core
> error: permission denied on key 'net.ipv4.route.flush'
> error: permission denied on key 'net.ipv6.route.flush'
> net.core.somaxconn = 128
> net.core.xfrm_aevent_etime = 10
> net.core.xfrm_aevent_rseqth = 2
> net.core.xfrm_larval_drop = 1
> net.core.xfrm_acq_expires = 30
> net.core.wmem_max = 131071
> net.core.rmem_max = 131071
> net.core.wmem_default = 126976
> net.core.rmem_default = 126976
> net.core.dev_weight = 64
> net.core.netdev_max_backlog = 1000
> net.core.netdev_tstamp_prequeue = 1
> net.core.message_cost = 5
> net.core.message_burst = 10
> net.core.optmem_max = 20480
> net.core.rps_sock_flow_entries = 0
> net.core.netdev_budget = 300
> net.core.warnings = 1
>
> I'm not very involved into zmq2 internals, but as i understood, socket
> buffer must have bigger size, that message_t class. And message_t has
> relatively small size than my net.core.wmem_default = 126976. May be
> zmq::mailbox_t::send tries to increase socket buffer in case when it
> have big size but is full too?
>

We don't have api plumbing to change mailbox socket buffer sizes yet,
ie. ZMQ_SNDBUF and ZMQ_RCVBUF don't apply.
You'd have to change them manually, can you try changing wmem_default
and rmem_default?
Note wmem_default and rmem_default are capped by wmem_max and
rmem_max, so you'd have to change them too.


Dhammika