Re: 2.5.70-mm9

[Andrew Morton]

Mingming Cao <[e-mail]> wrote:
>
> On Sat, 2003-06-14 at 01:01, Andrew Morton wrote:
> 
> > Was elevator=deadline observed to fail in earlier kernels?  If not then it
> > may be an anticipatory scheduler bug.  It certainly had all the appearances
> > of that.
> Yes, with elevator=deadline the many fsx tests failed on 2.5.70-mm5.
>  
> > So once you're really sure that elevator=deadline isn't going to fail,
> > could you please test elevator=as?
> > 
> Ok, the deadline test was run for 10 hours then I stopped it (for the
> elevator=as test).  
> 
> But the test on elevator=as (2.5.70-mm9 kernel) still failed, same
> problem.  Some fsx tests are sleeping on io_schedule().  
> 
> Next I think I will re-run test on elevator=deadline for 24 hours, to
> make sure the problem is really gone there.  After that maybe try a
> different Qlogic Driver, currently I am using the driver from Qlogic
> company(QLA2XXX V8).

Martin has just observed what appears to be the same failure on
2.5.71-mjb1, which is the deadline scheduler, using qlogicisp.

Again, some IO appears to have been submitted but it never came back.

It could be a bug in the requests queueing code somewhere, or in the device
driver.

So a good thing to do now would be to find the workload+IO
scheduler+filesystem which triggers it most easily, and run that with a
different device driver.  The feral driver (drivers/scsi/isp/ in -mm)
should be suitable for that test.