Rune Torgersen wrote:

Scott Wood wrote:

quoted

Almost certainly the latter.  Is the disk interrupt shared with any
other interrupts, that are marked IRQF_NODELAY?  The -rt
patch doesn't seem to handle mixing the two well.

Disk is on a muxed PCI interrupt. None of the other interrupts on the
mux is fireing at the time.

Regardless of whether they're firing, any request_irq with IRQF_NODELAY 
will turn off threading for all handlers.

Is is possible that the demuxer is not set up right? It is based loosely
on pq2-pci-pic.c

Try calling irq_set_chip_and_handler() with handle_level_irq, rather 
than irq_set_chip().  The -rt patch doesn't seem to have threadified the 
  __do_IRQ() path.

-Scott