[ECOS] cyg_selwakeup and performance impact?

["Zhichao Hong" <zhichao dot hong at gmail dot com>]

Hi,

One part of our application, the read thread cannot keep up with the
write for the serial driver when using pselect.  Since in the
serial.c, the cyg_selwakeup is used to tell the reader that is blocked
by a pselect when there is something available.  Sometimes, we found
that the thread will not get waken up when there are data to read.  A
workaround that we did is to comment out the condition in the function
serial_rcv_char() that wake up the thread in serial.c:
    //if( cbuf->nb == 0 )
        cyg_selwakeup( &cbuf->selinfo );

This change makes our application work as the reader will be waken up
more frequently from the pselect().  However, because cyg_selwakeup()
will __broadcast__ to all threads when this occurs.  This can be
expensive (sheduler lock/unlock, go through the queue and wake every
single thread).  Also, the serial.c is used as a kernel serial support
for our the customized serial drivers.  So changing here is probably
too pervasive.

I am wondering what other impact we may have after this change after
we tainted the eCos kernel.

Any comments and insights are much appreciated.

Zhichao


Zhichao Hong, CSDP
zhichao.hong@computer.org

-- 
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss