Red Hat Bugzilla – Bug 66143
System hang after 5-12 h IO stress - flushtlb problem?
Last modified: 2007-04-18 12:42:59 EDT
From Bugzilla Helper:
User-Agent: Mozilla/5.0 Galeon/1.2.0 (X11; Linux i686; U;) Gecko/20020408
Description of problem:
The problem occurs with all recent RedHat kernels. The symptopms seem to be less
severe with increasing kernel version (2.4.7-complete freeze, 2.4.9-31:
applications freeeze or die, 2.4.18-4: same symptoms, but system runs longer
before it happens).
Version-Release number of selected component (if applicable):
2.4.7-10, 2.4.9-31, 2.4.18-4
Steps to Reproduce:
1.Start high-load IO-bound stress test
(2.4.7) System appears "dead", with the exception of ping.
(2.4.9/2.4.18) Applications freeze or die. Alt-Sysrq still works.
Expected Results: Test completes successfully.
We have applied Sunil Saxenas's flushtlb patch:
and the problem appears to be gone.
Although that soesn't prove the flushtlb problem is actually causing our
problem, we strongly recommend that this patch be integrated
into the next RedHat update kernels.
That patch only affects SMP Pentium IV systems. Is the system in question also a
pentium IV one ?
Yes, it is a System GC LE chipset an Dual Prestonia.
In that case the yesterday released 2.4.9-34 kernel should fix this for 7.1/7.2;
for 7.3 a fix is in the works.
So you think that this is related to the PGE handling you mention
in the advisory?
In any case I have looked at the new kernel source - it is missing two small
Patches that were included in the 7.3 2.4.18 series. Both are very important for
our newer machines. I'll atttach them here although they are not directly
related to the problem itself.
Arjan, please have a look at them.
Concerning the original problem - we'll test the new kernel and see what happens.
Created attachment 59844 [details]
One-line patch for ServerWorks CSB5 IDE DMA
Created attachment 59845 [details]
4-line patch that fixes DMA address calculation >4GB (important!!)
Yes the PGE fix the the problem Sunil found.
As for the patches; I've added the 4Gb one to the tree in case we ever do a
2.4.9 erratum again.
I hope you do because that is really a nasty one if you have >4GB machines (we
had an Adaptec SCSI controller happily DMA'ing to and from the kernel core
memory). Meanwhile, we'll advise our >4GB customers to upgrade to the 7.3 kernel.
The CSB5-Patch may seem ridiculous - however, IDE load may cause such a heavy
interrupt load that timer and local APIC interrupts don't get through, causing
the LOC interupt counts to differ heavily, and (if the same CPU servers timer
and IDE IRQs) cause system time to go awry.
I am taking back what I said about the CSB5 patch. It should *not* be applied,
and probably even reverted in 2.4.18, until bug 66054 is resolved.
p4 bug is fixed in the 2.4.18-5 kernel