Re: timerfd incompatibility on mips

[Alexander Gordeev]

Ð Thu, 27 Jan 2011 17:49:39 -0800 (PST)
Davide Libenzi <davidel@xxxxxxxxxxxxxxx> ÐÐÑÐÑ:

> On Fri, 28 Jan 2011, Alexander Gordeev wrote:
> 
> > Ð Thu, 27 Jan 2011 16:53:39 -0800 (PST)
> > Davide Libenzi <davidel@xxxxxxxxxxxxxxx> ÐÐÑÐÑ:
> > 
> > > On Fri, 28 Jan 2011, Alexander Gordeev wrote:
> > > 
> > > > Hi!
> > > > 
> > > > I've just found a problem with TFD_NONBLOCK flag that can be passed to
> > > > timerfd_create.
> > > > 
> > > > <uclibc>/libc/sysdeps/linux/common/sys/timerfd.h which is installed
> > > > to /usr/include/sys/timerfd.h defines TFD_NONBLOCK as 04000 ie 0x800.
> > > > BTW glibc/eglibc headers do the same thing.
> > > > 
> > > > <linux>/include/linux/timerfd.h declares it as O_NONBLOCK which is
> > > > defined in fcntl.h. Usually O_NONBLOCK is 0x800 too but some
> > > > architectures including MIPS it redefine it:
> > > > 
> > > > arch/mips/include/asm/fcntl.h:#define O_NONBLOCK	0x0080
> > > > arch/alpha/include/asm/fcntl.h:#define O_NONBLOCK	00004
> > > > arch/sparc/include/asm/fcntl.h:#define O_NONBLOCK	0x4000
> > > > arch/parisc/include/asm/fcntl.h:#define O_NONBLOCK	000200004
> > > > 
> > > > My tests show that kernel thinks that TFD_NONBLOCK is 0x80 on MIPS. I
> > > > get what I want when I pass 0x80 and EINVAL when I pass 0x800. I don't
> > > > know why there are such uncertain things in Linux. Probably the problem
> > > > should be fixed in the kernel.
> > > > 
> > > > Right now one cannot use timerfd_create(..., TFD_NONBLOCK) on MIPS but
> > > > can use timerfd_create(..., O_NONBLOCK) instead.
> > > 
> > > Ther kernel definition (include/linux/timerfd.h) maps it directly to 
> > > O_NONBLOCK, but I am noticing that glibc bolt in the value.
> > > I don't think this is a kernel fix.
> > 
> > Yes, but linux/timerfd.h is not installed so timerfd.h from *libc
> > cannot include it. So adding it to include/linux/Kbuild would be a
> > first step to make them not hard-code the value. Right?
> 
> Glibc seems to wrap many files with their own definition.
> I am not sure why they are actually doing it, but they must have a valid 
> reason.
> It probably makes sense to just un-bolt the values from sys/timerfd.h and 
> link them to O_* values from fcntl.h.

Great! This is probably the simplest solution. Hope they'll use it.

However, I'm still curious why the value O_NONBLOCK value is
architecture-specific?

-- 
  Alexander

Attachment: signature.asc
Description: PGP signature