Problem: infinite hwm of version 4.2 lead to catch no msg

[laplaceyang]

In release version of 4.2, we got a change in zmq::pipe_t::set_hwms, that 0 means infinite hwm.

 void zmq::pipe_t::set_hwms (int inhwm_, int outhwm_)
 {
     int in = inhwm_ + inhwmboost;
     int out = outhwm_ + outhwmboost;
 
     // if either send or recv side has hwm <= 0 it means infinite so we should set hwms infinite
     if (inhwm_ <= 0 || inhwmboost <= 0)
         in = 0;
 
     if (outhwm_ <= 0 || outhwmboost <= 0)
         out = 0;
 
     lwm = compute_lwm(in);
     hwm = out;
 }

but in some case ,I use the dealer mode, and set sndhwm after connect, lead to the hwm and lwm eq 0. which is not my origin idea.

Then if one side (dealer mode)(side A) is infinite hwm, and another (router mode)(side B) set hwm to a limit number. Then A side will never send_activate_write to the peer, accoding to the follow code.

    if (lwm > 0 && msgs_read % lwm == 0)
        send_activate_write (peer, msgs_read);

Finally, we got a full hwm by check_hwm, and reset_pollin. Since then, no message can be caught.

[bluca]

This was changed to fix a bug, can't remember which, I'll dig it up later today

[bluca]

So I was thinking of this: #1649 but that's unrelated, I think.

Looking at the documentation, 0 meant infinite already in 2.1, so this hasn't really changed: http://api.zeromq.org/2-1:zmq-setsockopt

What changed is that now it's possible to modify the HWMs even after a bind/connect, since this PR: #1426 Is this what's causing you a problem?

Would you be able to upload to gist/pastebin a minimal test case that reproduces the issue you are facing? Thanks!

[laplaceyang]

@bluca I just write a test case:
https://gist.github.com/laplaceyang/40374e2b40a5bfec56c7ce69749418eb

In my test case, I write some code to simulate a normal service, which is that frontend svr send msg to backend svr, and wish to recv response from backend.

and in the code, frontend svr set hwm after connect, and send total 5000 msg (should bigger than default hwm 1000)

and if I build this code with version 4.1.6, I can get what I expected

then, I build with 4.2.0, I can only recv 1000 msg from backend.

[bluca]

The gist is not longer visible

[laplaceyang]

@bluca
https://gist.github.com/laplaceyang/40374e2b40a5bfec56c7ce69749418eb

here is the gist

I found something, that is
when we set hwm after connect, our hwm and lwm will be 0. because the outhwmboost and inhwmboost, which is prepared for inproc is default to be 0.

the pipe bind with socket_base and session_base is descibed as follow

socket_base  ( hwm 0, lwm 0 )
session_base  ( hwm 1000, lwm 500 )

session_base write the data to socket_base
but socket_base will never send_actived_write back to session_base
because of

    if (lwm > 0 && msgs_read % lwm == 0)
        send_activate_write (peer, msgs_read);

Thus, session_base is full.

So, I thank if we also set peer hwm when call set_hwm , will fix it.
Also the hwm/lwm should not always be 0 when use tcp and dealer and set hwm after connect.

I will try to fix it this week.

[laplaceyang]

@bluca I just wrote a thread-safe solution by using command ,and include your pr.

laplaceyang@3062a97

By simply test, it works. But afraid of include some mistake.

[bluca]

Thank you for taking care of it! I can have a look in detail later tonight

[bluca]

@laplaceyang that looks fine to me. Would you please rebase it and send a PR?

[laplaceyang]

@bluca pr has send, also with rebase my commit. #2302
Thankyou for your help

[bluca]

Thank you for fixing this! Closing as it's solved by PR #2302