Re: I finally prepared a testcase for read(inotify_fd) getting EINTR on PTRACE_ATTACH

[Oleg Nesterov <oleg@redhat.com>]

On 02/10, Denys Vlasenko wrote:
>
> On 02/10/2012 04:09 PM, Oleg Nesterov wrote:
>> On 02/10, Oleg Nesterov wrote:
>>> At first glance this looks obvious? I never used inotify and I never
>>> looked into fs/notify/inotify/, but it seems that inotify_read() simply
>>> returns -EINTR if signal_pending() and doesn't implement restarts.
>>>
>>> Probably this trivial change
>>>
>>> 	--- x/fs/notify/inotify/inotify_user.c
>>> 	+++ x/fs/notify/inotify/inotify_user.c
>>> 	@@ -264,7 +264,7 @@ static ssize_t inotify_read(struct file
>>> 			ret = -EAGAIN;
>>> 			if (file->f_flags&  O_NONBLOCK)
>>> 				break;
>>> 	-		ret = -EINTR;
>>> 	+		ret = -ERESTARTSYS;
>>> 			if (signal_pending(current))
>>> 				break;
>>>
>>>
>>> makes sense.
>>
>> except I meant -ERESTARTNOHAND to avoid the behavioural change.
>
> I run-tested the fix. It works: testcase no longer fails
> (modulo incorrect logic in the testcase which wase not working
> properly on "no bug detected" code path. Fixed one:
> http://sources.redhat.com/cgi-bin/cvsweb.cgi/~checkout~/tests/ptrace-tests/tests/eintr-on-attach.c?cvsroot=systemtap
> )

Good, thanks.

> I'm not sure inotify really wants to deviate from other reads
> and return -EINTR even for SA_RESTARTing signals. IOW:
> I think -ERESTARTSYS here would be more correct than -ERESTARTNOHAND.

I am not sure either. ERESTARTNOHAND doesn't change the behaviour,
that was my point.

But I agree, ERESTARTSYS makes more sense to me.

Oleg.