Re: 2.6.17-rc1-mm1

[Roger Luethi <rl@hellgate.ch>]

On Tue, 04 Apr 2006 15:09:53 -0700, Andrew Morton wrote:
> Zan Lynx <zlynx@acm.org> wrote:
> > Has anyone seen this yet?
> > 
> > BUG: scheduling while atomic: mii-tool/0x00000001/2968
> >  <c02db7f7> schedule+0x43/0x540   
> >  <c02dc617> schedule_timeout+0x7a/0x95
> >  <c011d687> process_timeout+0x0/0x5   
> >  <c011e8a4> msleep+0x1a/0x1f
> >  <e09100c9> rhine_disable_linkmon+0x40/0xf1 [via_rhine]   
[...]
> 
> hmm, according to git-whatchanged, this bug has been in there since October
> last year.  Weird that it hasn't been spotted before now.

It has been spotted [1] and diagnosed [2] about two weeks ago. I guess the
reason it took so long is that in addition to the debugging options, you
need a kernel that does call the spinlock code _and_ you need to look at
the kernel log even though the behavior is unchanged.

Anyhow, the mdelay to msleep conversion is useful only for a corner case: a
user with Rhine-I hardware (ancient) who needs low latency even while
fiddling with the NIC's media settings. I am not sure it's worth the
trouble. Any suggestions for an elegant solution?

Roger

[1] http://marc.theaimsgroup.com/?l=linux-netdev&m=114321570402396
[2] http://marc.theaimsgroup.com/?l=linux-netdev&m=114349201223976