Re: use O_DIRECT open file, when read will hang.

[Badari Pulavarty <pbadari@us.ibm.com>]

On Thursday 18 September 2003 10:54 pm, Andrew Morton wrote:
> Hugang <hugang@soulinfo.com> wrote:
> > Hello all:
> >
> > Steps to reproduce:
> >
> > rm -f /tmp/1.log
> > touch /tmp/1.log
> > echo << EOF > /tmp/hang.c
> > #include <sys/types.h>
> > #include <asm/fcntl.h>
> >
> > main()
> > {
> >         int i;
> >         char buf[1025];
> >
> >         i = open("/tmp/1.log", O_RDONLY | 040000, 0);
> >         if ( i != -1) {
> >                 read(i, buf, 1);
> >         }
> >         printf("'%s'", buf);
> > }
> > EOF
> > gcc -o /tmp/hang /tmp/hang.c
> > /tmp/hang
>
> This is due to O_DIRECT-race-fixes.patch forgetting to drop locks
> on error paths all over the place.
>
> I think this patch plugs them all for block-based direct-io, but it needs
> checking.
>
> There's also the little matter of (say) NFS direct-io which doesn't go
> through fs/direct-io.c at all; it will deadlock in a jiffy.
>
> Must say that I am getting very concerned about the general state of the IO
> paths in the -mm kernel.

Andrew,

I am also seeing some kind of regression on raw in 2.6.0-test5-mm2.
Unfortunately, this happens only with huge database benchmarks.
I still haven't narrowed it down.

So I am planning to do following to address your concerns.

1) I am going to do full code review of DIO & RAW code in -mm tree.

2) I want to test RAW & DIO code with various test cases we have.
	- fsx:
		on  RAW, DIO on files, AIO on RAW, AIO on files, AIO-DIO on files

	- rawiobench
		on  RAW, DIO on files, AIO on RAW, AIO on files, AIO-DIO on files

	- database benchmarks:
		on RAW, on files

	- can I do fsx, rawiobench on NFS files to test DIO on NFS ?

	- What else you would like to see ?

Thanks,
Badari