Minutes from LSE Con Call
A message that is not in the archive (sent off-list or not yet received): 29610000.1026499086@w-hlinder
Quoted 3 lines · Hanna Linder
8. Daniel asked about seeing results of rmap on a numaq. Martin got distracted but will get back to getting some results.
On elm3b17 I found more 2.5.25 vs. NUMA-Q issues and got workarounds for booting while doing work on stuffing pte_chains into highmem:
(1) cpu_init() heisenbug that goes away when adding printk's I just added the extra printk's and moved on
(2) in_interrupt() is racy and needs to disable preemption while checking per-cpu counters. The trivial fix.
(3) CONFIG_PREEMPT calls preempt_schedule() on secondaries before they're ready, and the exception is flagged as no vm86_info: BAD
Didn't want to go to deep into it at 3AM and disabled preempt.
The usual ioremap()/smp_call_function() deadlock, irqbalance plugging unreachable cpu's into ioapics, bio splitting to make qlogicisp happy, and MAX_IO_APICS things are still needed.
So I guess the "real fixes" need to get found at some point and propagated out to maintainers. hpa and mochel are on cpu_init(), and rml should be on preempt vs. printk and in_interrupt(). axboe has already got bio splitting ready and no idea about the MAX_IO_APICS and getting irqbalance fixes past mingo doesn't look easy.
Cheers,
Bill
Quoted 4 lines · William Lee Irwin III
(3)
CONFIG_PREEMPTcallspreempt_schedule()on secondaries before they're ready, and the exception is flagged as novm86_info: BAD
Didn't want to go to deep into it at 3AM and disabled preempt.
Sorry, I forgot to mention that preempt_schedule() was called from printk(), specifically, release_console_sem().
Cheers,
Bill