Re: [Xenomai-help] writing from user space through fifo problems

[Philippe Gerum <philippe.gerum@domain.hid>]

Gilles Chanteperdrix wrote:
> On Fri, Feb 22, 2008 at 3:45 AM, NZG <metalninjadragon@domain.hid> wrote:
>> On Thursday 21 February 2008 10:23:15 am Jan Kiszka wrote:
>>  > That's good. This test will also allow you to check it against
>>  >  - more recent Xenomai (please check the Changelog for pipe-related
>>  >    changes since 2.3.1, I bet there are a few)
>>  >  - some other arch, specifically x86 (IIRC, there is still some Linux
>>  >    scheduling artifact left in the I-pipe patch for blackfin, but it
>>  >    shouldn't make a difference here - well, who knows...)
>>  Well, in the process of writing a simple test I kindof found the problem.
>>  My user space program is writing and then immediatly closing, which appears to
>>  destroy the data if it hasn't been read by then. Inserting a sleep between
>>  write and close causes it to work.
>>
>>  I noticed this because my test program was working because it just so happened
>>  to pause until the real time task ended before closing the fd.
>>
>>  I remember doing this before, now that it's figured out, but it's still kind
>>  of problematic because it doesn't appear to be possible to block on user
>>  space writes.
>>
>>  Opening with O_SYNC does not cause it to block, and fsync(fd) doesn't seem to
>>  do anything. anybody have any suggestions on how to know data has been
>>  received before closing a pipe after a write? Surely somebody's done this
>>  before.
> 
> Well, I do not know the details of the rt_pipe implementation, but
> intuitively I would say that what you observe is not normal: when you
> write to the rt_pipe, the reader should be woken up immediately and
> consume the data, before the producer can close the descriptor. Are
> you writing data to the pipe from an RT task ?
> 

As far as I understand, the reader polls, so userland may close the
fildes before the real-time task peeks at its pipe end.

-- 
Philippe.