Re: ionice kills vanilla 2.6.0-test9 was [Re: [PATCH] cfq + io priorities (fwd)]

[Linus Torvalds]

[ Trivial test input file included, and "bug-binutils" added to the CC ]

On Fri, 21 Nov 2003, Linus Torvalds wrote:
> 
> Just do "gdb vmlinux" and do "disassemble system_call" to see this. It 
> says something like
> 
>       cmp    $0x448,%eax
>       jae    syscall_badsys
> 
> and it _should_ use just 274 (0x112 rather than 0x448) on x86.
> 
> Now, I wonder if this has ever worked, of which binutils version broke 
> it..

Some looking around shows that it works correctly on a SuSE box with

        GNU assembler version 2.13.90.0.18 (i486-suse-linux) using BFD version 
2.13.90.0.18 20030121 (SuSE Linux)
        GNU ld version 2.13.90.0.18 20030121 (SuSE Linux)

but is broken on RH Fedora core 1:

        GNU assembler version 2.14.90.0.6 (i386-redhat-linux) using BFD version 
2.14.90.0.6 20030820
        GNU ld version 2.14.90.0.6 20030820

So it definitely _does_ work in some versions, and the bug appears to be 
new to binutils 2.14, with 2.13 doing the right thing.

You can trivially see if with a simple assembly file like

        start:
                .long 1,2,3,a
        a=(.-start)/4

where 2.13.90 as shipped by SuSE will get it right (and generate a list of
1,2,3,4), while 2.14.90 from Fedora core will generate 1,2,3,16.

                Linus