Robert Watson wrote: > I see the same problem in 4.x. Drew and I exchanged some e-mails > comiserating over the fact that this seems to be the case on a lot of > machines. Not that this is necessarily what you're seeing, > but it would > be interesting to know whether in previous configurations you see > different symptoms of a similar problem. With the old configuration (a kernel from 3 or 4 days ago, ACPI disabled (but APIC enabled since it's an SMP system), and two of the PCI cards switched around), this is what I was getting: (copied from the "Current crash on today's kernel" thread) # vmstat -i interrupt total rate irq1: atkbd0 2 0 irq0: clk 5142796 995 irq4: sio0 1056 0 irq6: fdc0 10 0 irq8: rtc 658343 127 irq13: npx0 1 0 irq14: ata0 205255 39 irq15: ata1 205046 39 irq16: em0 atapci1 11635073 2251 irq17: em1 ahc0++ 6549430 1267 irq19: atapci4+ 442467 85 Total 24839479 4807 As you can see, em0 (the source with the highest interrupt load) didn't sit on its own ithread. Instead it shared it with the Promise SATA150 TX4 controller. em1/ahc0/atapci2+ shared irq17, and the interrupt count on that ithread looks about right. Something in this configuration has crashed my machine 3 times in 6 days, so I'm not looking to go back unless the new config also results in crashes. Back to the current config (ACPI enabled): What about running the kernel with DEVICE_POLLING enabled? I've seen posts here saying it doesn't make any sense to do so in the SMP case, but I've also seen people report that they are doing it (by modding the appropriate file to allow the compile to finish). Does it really not make any sense doing it in SMP, or is it a matter of insufficient testing/debugging of the code, or are there other reasons not to do it? My thinking here is that if the NIC doesn't generate any interrupts, the ithread that reports the "ghost" interrupts (atapci1+) should no longer do that. Maybe the real question is: does the ithread that reports the "ghost" interrupts actually service all the interrupts, or is 'vmstat -i' reporting something that really doesn't consume any CPU time? /Daniel ErikssonReceived on Tue Oct 19 2004 - 13:46:37 UTC
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:38:18 UTC