[2.5.74] bad: scheduling while atomic!

I haven't had the time to investigate this, so I don't have much information to share beyond the trace below. I think I have seen this at least with 2.5.73, too. The system looks okay, then, usually hours later (if at all, it's a rare event), something triggers a flood of those call traces (many of them per second).

The syslog seems to suggest it might be related to IDE DMA:

Jul  4 17:17:28 [kernel] hda: dma_timer_expiry: dma status == 0x61
Jul  4 17:17:44 [kernel] hda: timeout waiting for DMA
Jul  4 17:17:44 [kernel]  [<c0107000>] default_idle+0x0/0x40
Jul  4 17:17:44 [kernel] bad: scheduling while atomic!

Compiler is gcc 3.2.3.

bad: scheduling while atomic!

Call Trace:
 [<c0107000>] default_idle+0x0/0x40
 [<c011f110>] schedule+0x500/0x510
 [<c0107063>] poll_idle+0x23/0x40
 [<c0118073>] apm_cpu_idle+0xa3/0x140
 [<c0117fd0>] apm_cpu_idle+0x0/0x140
 [<c0107000>] default_idle+0x0/0x40
 [<c01070b8>] cpu_idle+0x38/0x40
 [<c0105000>] rest_init+0x0/0x30
 [<c037c738>] start_kernel+0x138/0x140
 [<c037c4c0>] unknown_bootoption+0x0/0x100

Possibly the IDE error handler has a locking imbalance. It returned from the interrupt handler without having unlocked a lock which it should have unlocked, and that left the currently-running process (the idle task in this case) with an incorrect preempt count.