silo overflows

[Matthew Dillon]

:Andrew Atrens wrote:
:
:> Seen on my home system ( 56k modem ) -
:> 
:> Sep  9 22:35:49 atrens-home /kernel: sio0: 7 more silo overflows (total 35)
:> 
:> these seem to be inert but are happening quite a lot - essentially 
:> whenever I'm transferring data (telnet, ftp, web) through the modem.
:
:	Andrew,
:
:	I was browsing through the GNATS database and I remembered this
:	post.  You might wanna checkout the following PRs, which seem to
:	be reporting bugs to do with silo overflows.  One of the PRs
:	also has a workaround, which you might want to try and tell us
:	if it worked.
:
:	PR (1): http://www.freebsd.org/cgi/query-pr.cgi?pr=i386/41569
:	PR (2): http://www.freebsd.org/cgi/query-pr.cgi?pr=i386/26261
:
:	Regards,
:
:-- 
:Hiten Pandya
:hmp at xxxxxxxxxxxxx

    There are a couple of issues here.    First, until we unwind the
    big giant lock and/or remove it from the fast interrupt path, there
    are going to be latency issues with DragonFly's servicing of interrupts
    and characters may be lost even if the serial hardware FIFO level is
    set to FIFO_RX_LOW.

    The hardware fifo is typically 16 characters.  FIFO_RX_HIGH interrupts
    when there are 4 characters left.  FIFO_RX_MEDH interrupts when there
    are 8 characters left (I think).  Normally this would be sufficient
    time for an interrupt to come and clean out the FIFO (8 chars @ 115200
    == 694 uS, but due to the current issues latencies can be much higher.

    Lower FIFO levels will reduce the silo errors but they will also
    be severely inefficient from a cpu perspective.  You might remember
    that in FreeBSD-4  Bruce had set the level to FIFO_RX_HIGH and I
    rammed through a change to default it to FIFO_RX_MEDHI at baud 
    rates higher then 4800 bps.

    I don't think we want to lower the FIFO level any more then MEDHI,
    at least not until we clean up the interrupt path.  I recommend
    reducing the host<->modem baud rate from 115200 (which is what it is
    likely set at) to 57600.

					-Matt
					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>