On Mon, Apr 18, 2005 at 01:32:46PM +0400, Pavel Kiryukhin wrote:
> Hi,
> the preempt_disable/preempt_enable sequence in do_cpu() [traps.c]
> exists quite long (patch submitted in Oct. 2004), so it should be nothing
> wrong there.
>
> Can somebody please comment why use of preempt_disable/enable in do_cpu
> will not result in "scheduling while atomic" for fpu-less cpu (with enabled
> preemption).
>
> The sequence looks like
>
> do_cpu()
> | preempt_disable()
> | fpu_emulator_cop1Handler()
> | | cond_reshed()
> | | | schedule() <------ scheduling while atomic
>
>
> The proposed patch was tested for Sibyte, but it has fpu (AFAIK) and has no
> fpu_emulator_cop1Handler called.
>
fpu_emulator maintains global variables and in general is dangerous
to be preempted in the middle of processing.
The quick fix for this problem is probably to move preemption disabling/
enabling inside fpu_emulator_cop1Handler().
Better fix is probably to modify fpu emulator so that it is preemption
safe overall.
Jun
|