Re: pull request: SCTP updates for net-next

[Vlad Yasevich <vladislav.yasevich@hp.com>]

Andrew Morton wrote:
> On Mon, 23 Nov 2009 16:06:50 -0500 Vlad Yasevich <vladislav.yasevich@hp.com> wrote:
> 
>> Andrei Pelinescu-Onciul (3):
>>       sctp: allow setting path_maxrxt independent of SPP_PMTUD_ENABLE
>>       sctp: limit maximum autoclose setsockopt value
>>       sctp: fix integer overflow when setting the autoclose timer
> 
> Problems with this one:
> 
> : commit f6778aab6ccc4b510b4dcfa770d9949b696b4545
> : Author:     Andrei Pelinescu-Onciul <andrei@iptel.org>
> : AuthorDate: Mon Nov 23 15:54:01 2009 -0500
> : Commit:     Vlad Yasevich <vladislav.yasevich@hp.com>
> : CommitDate: Mon Nov 23 15:54:01 2009 -0500
> : 
> :     sctp: limit maximum autoclose setsockopt value
> :     
> :     To avoid overflowing the maximum timer interval when transforming
> :     the  autoclose interval from seconds to jiffies, limit the maximum
> :     autoclose value to MAX_SCHEDULE_TIMEOUT/HZ.
> :     
> :     Signed-off-by: Andrei Pelinescu-Onciul <andrei@iptel.org>
> :     Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
> : 
> : diff --git a/net/sctp/socket.c b/net/sctp/socket.c
> : index d2681a6..71513b3 100644
> : --- a/net/sctp/socket.c
> : +++ b/net/sctp/socket.c
> : @@ -2086,6 +2086,9 @@ static int sctp_setsockopt_autoclose(struct sock *sk, char __user *optval,
> :  		return -EINVAL;
> :  	if (copy_from_user(&sp->autoclose, optval, optlen))
> :  		return -EFAULT;
> : +	/* make sure it won't exceed MAX_SCHEDULE_TIMEOUT */
> : +	if (sp->autoclose > (MAX_SCHEDULE_TIMEOUT / HZ) )
> : +		sp->autoclose = MAX_SCHEDULE_TIMEOUT / HZ ;
> :  
> :  	return 0;
> :  }
> 
> a) it has two coding-style errors in two lines.  Please go away, add
>    scripts/checkpatch.pl to your patch development tools and then continue
>    reading.
> 
> b) have you done that yet?

Gack, totally missed those spaces in the email client.  Will use checkpatch from
now on.

> 
> c) it generates this on 64-bit:
> 
> net/sctp/socket.c: In function 'sctp_setsockopt_autoclose':
> net/sctp/socket.c:2090: warning: comparison is always false due to limited range of data type
> 
>    but that's proving somewhat hard to fix in a nice way.

There was a patch that fixed it.  I saw that you pulled it into -mm tree.

> 
> d) I'm not sure that we should fix it anyway.  Is it really a good
>    idea to take an incorrect, invalid setting from userspace, to
>    silently modify that setting and to not inform userspace?
> 

We do it with other options, that's not an issue.  At the values where
we change it, it becomes somewhat moot point.

>    Bear in mind that MAX_SCHEDULE_TIMEOUT has different values on
>    32- and 64-bit kernels.  So the same source code will have different
>    behaviour depending on what type of kernel it is executed on.
> 
>    I think.
> 
>    It also means that kernel behaviour will differ as CONFIG_HZ is
>    altered, in some way which I can't be bothered working out.
> 

This is actually OK.  It allows the longest possible timeout depending
on configuration.  I've actually tested this in a harness and it works
very well with different HZ values (to bad the harness in not 64 bit aware. :( )

> 
> Overall, it would be way simpler and saner to clamp this value to some
> explicit time period, IMO.
> 
> <pulls number out of thin air>
> 
> --- a/net/sctp/socket.c~a
> +++ a/net/sctp/socket.c
> @@ -2086,9 +2086,8 @@ static int sctp_setsockopt_autoclose(str
>  		return -EINVAL;
>  	if (copy_from_user(&sp->autoclose, optval, optlen))
>  		return -EFAULT;
> -	/* make sure it won't exceed MAX_SCHEDULE_TIMEOUT */
> -	if (sp->autoclose > (MAX_SCHEDULE_TIMEOUT / HZ) )
> -		sp->autoclose = MAX_SCHEDULE_TIMEOUT / HZ ;
> +	/* make sure it won't exceed one hour */
> +	sp->autoclose = min_t(u32, sp->autoclose, 60 * 60);
>  

But that may not be long enough.  The spec doesn't impose limits
and it's really up to the application to decide how long it wants
to keep idle connections open.  Thus any limits shorter the maximum
supported by kernel are really artificial and may not be sufficient.

Thanks
-vlad