Messages in this thread | | | Date | Fri, 18 Sep 2009 11:46:42 +0200 | From | Ingo Molnar <> | Subject | Re: [Patch 1/2] ia32: use generic sys_pipe() |
| |
* Al Viro <viro@ZenIV.linux.org.uk> wrote:
> On Tue, Aug 04, 2009 at 02:24:32PM +0200, Ingo Molnar wrote: > > > Please _ALWAYS_ mention the change in behavior in the changelog, > > just in case someone ends up bisecting it. I only found out when i > > reviewed the two syscalls out of caution. > > > > The syscall you remove kept stale fd's around in case of -EFAULT > > from copy_to_user(). The generic version does an explicit close of > > those files: > > > > sys_close(fd[0]); > > sys_close(fd[1]); > > error = -EFAULT; > > > The generic version looks like the better choice to me but this > > difference should be mentioned in the changelog nevertheless, just > > in case some buggy app runs into this issue. > > It's not a matter of QOI, actually - sys32_pipe() is supposed to do > what sys_pipe() would do on i386 host. So any difference in handling > of an error case is simply wrong. > > Whether we want those sys_close() in sys_pipe() or not is a separate > question, but we definitely want the same behaviour when 32bit process > is run natively and when it's run on amd64. So sys32_pipe() has no > business existing at all.
I generally agree except the notion that we 'must' do this change.
This is a (small) past ABI sin of ours and we have to live with the consequences without breaking apps. We didnt break any it appears so it's all good.
Ingo
| |