Sv: Sv: Sv: Sv: Sv: Sv: Sv: Named pipes and multiple writers

[Ken Brown]

On 4/1/2020 1:14 PM, sten.kristian.ivarsson@gmail.com wrote:
>> On 4/1/2020 4:52 AM, sten.kristian.ivarsson@gmail.com wrote:
>>>> On 3/31/2020 5:10 PM, sten.kristian.ivarsson@gmail.com wrote:
>>>>>> On 3/28/2020 10:19 PM, Ken Brown via Cygwin wrote:
>>>>>>> On 3/28/2020 11:43 AM, Ken Brown via Cygwin wrote:
>>>>>>>> On 3/28/2020 8:10 AM, sten.kristian.ivarsson@gmail.com wrote:
>>>>>>>>>> On 3/27/2020 10:53 AM, sten.kristian.ivarsson@gmail.com wrote:
>>>>>>>>>>>> On 3/26/2020 7:19 PM, Ken Brown via Cygwin wrote:
>>>>>>>>>>>>> On 3/26/2020 6:39 PM, Ken Brown via Cygwin wrote:
>>>>>>>>>>>>>> On 3/26/2020 6:01 PM, sten.kristian.ivarsson@gmail.com wrote:
>>>>>>>>>>>>>>> The ENIXIO occurs when parallel child-processes
>>>>>>>>>>>>>>> simultaneously using O_NONBLOCK opening the descriptor.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> This is consistent with my guess that the error is
>>>>>>>>>>>>>> generated by fhandler_fifo::wait.  I have a feeling that
>>>>>>>>>>>>>> read_ready should have been created as a manual-reset
>>>>>>>>>>>>>> event, and that more care is needed to make sure it's set
> when it should be.
> 
> [snip]
> 
>>>>>>> Never mind.  I was able to reproduce the problem and find the cause.
>>>>>>> What happens is that when the first subprocess exits,
>>>>>>> fhandler_fifo::close resets read_ready.  That causes the second
>>>>>>> and subsequent subprocesses to think that there's no reader open,
>>>>>>> so their attempts to open a writer with O_NONBLOCK fail with ENXIO.
> 
> [snip]
> 
>>> I wrote in a previous mail in this topic that it seemed to work fine
>>> for me as well, but when I bumped up the numbers of writers and/or the
>>> number of messages (e.g. 25/25) it starts to fail again
> 
> [snip]
> 
>> Yes, it is a resource issue.  There is a limit on the number of writers
> that can be open at one
>> time, currently 64.  I chose that number arbitrarily, with no idea what
> might actually be
>> needed in practice, and it can easily be changed.
> 
> Does it have to be a limit at all ? We would rather see that the application
> decide how much resources it would like to use. In our particular case there
> will be a process-manager with an incoming pipe that possible several
> thousands of processes will write to

I agree.

> Just for fiddling around (to figure out if this is the limit that make other
> things work a bit odd), where's this 64 limit defined now ?

It's MAX_CLIENTS, defined in fhandler.h.  But there seem to be other resource 
issues also; simply increasing MAX_CLIENTS doesn't solve the problem.  I think 
there are also problems with the number of threads, for example.  Each time your 
program forks, the subprocess inherits the rfd file descriptor and its 
"fifo_reader_thread" starts up.  This is unnecessary for your application, so I 
tried disabling it (in fhandler_fifo::fixup_after_fork), just as an experiment.

But then I ran into some deadlocks, suggesting that one of the locks I'm using 
isn't robust enough.  So I've got a lot of things to work on.

>> In addition, a writer isn't recognized as closed until a reader tries to
> read and gets an error.
>> In your example with 25/25, the list of writers quickly gets to 64 before
> the parent ever tries
>> to read.
> 
> That explains the behaviour, but should there be some error returned from
> open/write (maybe it is but I'm missing it) ?

The error is discovered in add_client_handler, called from thread_func.  I think 
you'll only see it if you run the program under strace.  I'll see if I can find 
a way to report it.  Currently, there's a retry loop in fhandler_fifo::open when 
a writer tries to open, and I think I need to limit the number of retries and 
then error out.

Ken