linuxsh-dev

[linuxsh-dev] floating point usage in SH4 linux

[Pete L. <pet...@st...>]

Hi all,
I have a question concerning floating point support in the Linux kernel for
SH4 processors.

My situation is this, I am trying to port an x86 driver to an SH4 HARP
board.  The x86 driver makes extensive use of floating point math, which I
know is not supposed to be used in Linux Kernel mode, but I do not have an
alternative approach to avoid making use of floating point.

My question is this, is there a version of kernel, or the tool chain that is
more likely to support the usage of floating point in the kernel?
Alternatively, is it possible to say definitively that floating point math
will or will not work in a kernel mode driver?

Thank you for your time, it is much appreciated,
- pete

Re: [linuxsh-dev] floating point usage in SH4 linux

[David W. <dw...@in...>]

If you think you want FP in the kernel, you're almost certainly approaching 
the problem wrong. 

Why can't you move the FP tasks to userspace?

--
dwmw2

[linuxsh-dev] floating point usage in SH4 linux

[NIIBE Y. <gn...@m1...>]

Current 2.5.3-pre5 (or our CVS) doesn't support FPU usage in kernel.
Once I've tried to support it (and I think it worked well at that
time), but eventually, I gave up.

You can check, say, 2.4.0 implementation, if you really have
interests.

It goes like this:
    We have FPU lazy switching mechanism.
    Wnen kernel asks FPU use, we pretend as if the task "init" use FPU.
    On Kernel -> Kernel transition, we save FPU registers on kernel stack.

This works well _provided_ the use of FPU in kernel can be seen as if
"init" use FPU.

However, the kernel thread softirqd has been introduced in 2.4.X.
With it, the bottom-half of the driver is invoked by that thread.  If
bottom-half uses FPU, it results quite a mess, because it breaks
the condition (of it's "init" who use FPU).
--