user-space hangs - followup
- From: Dan Miller <dan@xxxxxxxxxxxxxxx>
- Date: Mon, 07 Aug 2006 13:52:41 -0500
I had previously posted regarding a problem that I'm seeing here. The
characteristics of the problem are that all kernel threads and interrupts
appear to continue working, but all user-space processes appear to hang.
Symptoms of the "hang" are that network traffic which passes THROUGH the
box continue to pass, and pings to the box succeed, but any tcp/udp
traffic to or from the box will fail during the hang. Another quirk of
the situation is that the hang is not permanent, but will resolve itself
after a period of time (anywhere from 15 minutes to 8 hours).
Gil Hammond (GH) had suggested (on this newsgroup) that it sounded as
though all user processes were getting stacked up behind a lock/mutex or
other resource in the kernel; I've been trying to pursue this line of
reasoning further to research this problem (which is still occuring).
Another engineer that I spoke with elsewhere, observed that it didn't
sound like a simple mutex deadlock-type issue, because a deadlock like
that should never recover. He observed that it sounded more like a
counter that was wrapping or otherwise going out of range. Based on that
theory, I've been looking back through my driver again - a good exercise
anyway, since I've been working on this driver for eight years now, and
there are some parts that I don't remember much anymore!!
One thing I'm finding is that we have code such as the following in many
parts of the driver:
timeout_jiffies = jiffies + jiffies_from_ms(timeout_msec) ;
while ( (rc = test_some_other_condition(device)) &&
(jiffies < timeout_jiffies ) ) {
schedule();
}
// Did we succeed in getting the MUTEX flag?
if ( rc != 0 ) {
return -1;
}
return 0;
I now note that this is actually vulnerable to counter-wrapping; for
example, if timeout_jiffies gets set to (MAX_JIFFIES - 5), and
test_some_other_condition() takes more than 5 jiffies, the counter would
wrap and we'd wait awhile for the timeout to occur. (mind you, we have
an 850Mhz Celeron in our machines, so wrapping should still only take
about 8.4 minutes). Anyway, my question at this point is this:
Although the counter could wrap and leave me in this loop for awhile, I'm
still calling schedule(), as I should. Now this will allow at least
kernel-space tasks to proceed, but what about user-space processes??
Would they end up blocked by this loop?? Or will they still get
scheduled??
Dan Miller
----== Posted via Newsfeeds.Com - Unlimited-Unrestricted-Secure Usenet News==----
http://www.newsfeeds.com The #1 Newsgroup Service in the World! 120,000+ Newsgroups
----= East and West-Coast Server Farms - Total Privacy via Encryption =----
.
- Follow-Ups:
- Re: user-space hangs - followup
- From: gil_hamilton@xxxxxxxxxxx
- Re: user-space hangs - followup
- Prev by Date: Re: gcc optimization changes results
- Next by Date: Re: gcc optimization changes results
- Previous by thread: Load balancing using SCHED_RR
- Next by thread: Re: user-space hangs - followup
- Index(es):
Relevant Pages
|
